Catena 118 (2014) 124–135

Contents lists available at ScienceDirect

Catena
journal homepage: www.elsevier.com/locate/catena

Earthquake induced landslide susceptibility mapping using an integrated

ensemble frequency ratio and logistic regression models in West
Sumatera Province, Indonesia
Zahrul Umar a, Biswajeet Pradhan b,⁎, Anuar Ahmad a, Mustafa Neamah Jebur b, Mahyat Shafapour Tehrany b
a
Department of Geomatic Engineering, Faculty of Geoinformation and Real Estate, Universiti Teknologi Malaysia, Skudai, Johor, Malaysia
b
Department of Civil Engineering, Faculty of Engineering, Universiti Putra Malaysia (UPM), 43400 Serdang, Selangor, Malaysia

a r t i c l e i n f o a b s t r a c t

Article history: An 8 Richter Scale (RS) earthquake struck West Sumatra on Wednesday, 30 September 2009, at 17.16 pm which
Received 6 November 2013 led to huge number of landslides. Hence a comprehensive landslide susceptibility mapping (LSM) should be
Received in revised form 21 January 2014 produced in order to reduce the damages to people and infrastructures. In the international landslide literature,
Accepted 17 February 2014
various statistical methods such as frequency ratio (FR) and logistic regression (LR) have been widely used indi-
Available online 7 March 2014
vidually for LSM, but they have some weaknesses. FR which is able to perform bivariate statistical analysis (BSA)
Keywords:
assesses the influence of classes of each conditioning factor on landslide occurrence. However, the correlation
Landslide between the factors is mostly neglected. On the other hand, LR is able to analyze the relationship among the fac-
Earthquake tors while it is not capable to evaluate the classes of each landslide conditioning factor. This paper aims to propose
Ensemble an ensemble method of FR and LR in order to overcome their weak points. For LSM, a landslide inventory map
Geographic information system (GIS) with a total of 87 landslide locations was extracted from various sources. Then the landslide inventory was
Remote sensing (RS) randomly divided into two datasets 70% for training the models and the remaining 30% was used for validation
Indonesia purpose. The landslide conditioning factors consist of: altitude, curvature, river, SPI, rainfall, soil type, soil texture,
land use/cover (LULC), peak ground acceleration (PGA), geology, slope, aspect, lineament and topographic
wetness index (TWI). Four PGA of 7.5, 8, 8.6 and 9 were acquired and PGA 8 which was related to the 2009 earth-
quake was used to generate the model. Finally, the produced landslide susceptibility maps were validated using
an area under the (ROC) curve method. For the model which was derived by PGA 8, the validation results showed
84% and 78% success and prediction rates respectively. Furthermore, the prediction rates for the models made by
PGA 7.2, 8.6 and 9 are 79%, 78% and 81% respectively. The result proved the reasonable efficiency of the proposed
method for earthquake induced landslide susceptibility mapping. Also the proposed ensemble method can be
used in other hazard studies as it is capable to produce rapid and accurate assessment for disaster management
and decision making.
© 2014 Elsevier B.V. All rights reserved.

1. Introduction steep slopes (Karig, 1974). Volcanic eruption of the material has high
porosity, less compact and scattered in an area with a steep slope. If
Landslide is one of the dangerous natural hazards and its frequency this condition disturbs the hydrological balance, the area will be prone
of occurrence is increasing worldwide (Jebur et al., 2013a; Pradhan and to landslides (Van Asch et al., 1999). The territory of Indonesia in gener-
Lee, 2010b; Yin et al., 2010). Landslides can cause damages to transpor- al and the Province of West Sumatra in particular are mostly hills and
tation, people and properties (Mohammady et al., 2012). The mountains which are susceptible to landslides. Frequent and high
Indonesian archipelago sits on the junction of the world's three great intensity precipitation, and earthquakes are naturally triggering factors
plates namely the Indo-Australian plate, the Eurasia plate, and the in landslide occurrence. Based on the review of the locations of land-
Pacific plate. The junction is situated in the subduction zone which slides in Padang, the locations of landslides were estimated to be in
resulted in the formation of a volcanic island arc with moderate to the line of the fault. The hills are composed of tuft and pumice rides
above andesitic rocks with 20 m thickness which made this area very
prone to landslides (Leo et al., 1980).
An 8 Richter Scale (RS) earthquake struck West Sumatra and its
⁎ Corresponding author. Tel.: +60 3 89466383; fax: +60 3 89466459. surrounding areas on Wednesday, 30 September 2009, at 17.16 pm.
E-mail addresses: [email protected], [email protected] (B. Pradhan). The epicenter of this earthquake was located at latitude 0° 84′ South,

http://dx.doi.org/10.1016/j.catena.2014.02.005
0341-8162/© 2014 Elsevier B.V. All rights reserved.
 Z. Umar et al. / Catena 118 (2014) 124–135 125

and longitude 99° 65′ East, 57 miles South-West of the city of Pariaman, the suitable and non-suitable areas for development activities (Gupta
with a depth of 71 km (Indonesia Meteorologi and Geophysics Body et al., 2008; Jebur et al., 2013b; Saha et al., 2005; Sarkar and Kanungo,
(BMKG), 2009). The earthquake destroyed many districts/towns in 2004; Van Westen et al., 2003, 2008). Many methods and techniques
this area and caused large number of deaths (Satkorlak Disaster of have been proposed to evaluate landslide prone area using these two
West Sumatra, 2009). In addition to the Semangko fault and the techniques (Konadu and Fosu, 2009; Lee and Pradhan, 2007; Pradhan,
meeting point between the two tectonic plates, the land of West 2010; Pradhan and Lee, 2010a,b).
Sumatra is also composed of intermingling highlands; starting from The most popular methods for landslide hazard mapping are
the Province of Aceh in the north to the Province of Lampung in the deterministic approach (or safety factor) (Westen and Terlien, 1996),
south (Islam and Khan, 1986). Some of the volcanic mountains on this heuristic approach (Barredo et al., 2000; Van Westen, 2000), statistical
highland are still active. The main rock types in this highland are also approach (Luzi et al., 2000), support vector machine (SVM)
truncated by fissures, and cracks. They are destroyed by earthquakes (Pourghasemi et al., 2013; Wan and Lei, 2009), neuro-fuzzy based
and this led to the occurrence of many landslides in the region study (Akgun et al., 2012; Oh and Pradhan, 2011), fuzzy logic
(Wilkinson et al., 2012). For this reason, it is important to detect the sus- (Pradhan, 2011), and artificial neural network (ANN) (Chi et al., 2002).
ceptible areas in order to predict future landslides. Landslide susceptible Although these methods are capable to recognize the susceptible areas
areas can be assessed and predicted through scientific analysis and thus and produce the landslide susceptibility map, they have some disadvan-
landslide damages can be decreased through the proper prevention ac- tages that reduce the efficiency of the predictive models when applied
tions (Ozdemir and Altural, 2012; Pradhan, 2013). Over the last decades, individually (Tien bu et al., 2012c,d). Some comparative studies have
numerous efforts have been devoted by many researchers to develop been done in order to compare these methods and analyze their perfor-
landslide susceptibility maps worldwide using GIS (Lee and Pradhan, mance such as Pradhan (2010) and Pradhan and Lee (2010a). In a recent
2007; Pourghasemi et al., 2013; Pradhan and Lee, 2010a; Pradhan, paper, Pradhan (2013) used decision tree (DT), SVM and adaptive neuro-
2013; Bui et al., 2013; Pradhan et al., 2010c, 2011). fuzzy inference system (ANFIS) for LSM of Penang Hill, Malaysia in order
The impact of earthquake on landslide occurrence has been reported to compare their efficiency. Defining the rules for DT and selecting the
by many studies (Bai et al., 2012; Sidle and Ochiai, 2006; Zhou et al., SVM parameters are hard task and it is very time consuming. Although
2013). Yin et al. (2009) analyzed the distribution of landslides triggered ANFIS performs better than the others, it entails a large number of pa-
by the earthquake in Wenchuan County, Sichuan Province, China on 12 rameters. All three methods require a very high speed computer which
May 2008. About 20,000 fatalities have been reported due to the impact can handle heavy analysis (Chau et al., 2005).
of 15,000 landslides which was made by the Wenchuan earthquake. In a The ANN model is one the most popular methods in many fields and
recent paper, Oh et al. (2010) evaluated the impact of earthquakes especially in landslide analysis (Wan et al., 2010b). However, the ANN
which is often aggravated by the triggering of landslides. They could model is considered as a black box which has complex process and its
discover that the areas affected by landslides are similar to other parts performance is not easy to understand (Pradhan and Buchroithner,
of the world and the size of the earthquakes has a direct effect on the 2010). Also this method cannot produce accurate predictions in the
size of landslides. Probabilistic approaches are based on the observed case that the validation dataset contains values outside the range of
relationships between each conditioning factor and the distribution of those used for training. Also in the case that the large number of
landslides. Remote sensing (RS) and geographic information system variables is used, it makes the entire modeling process time consuming
(GIS) are efficient techniques for LSM and they are capable to identify (Ghalkhani et al., 2012). On the other hand, in the qualitative

Fig. 1. Study area map.

126 Z. Umar et al. / Catena 118 (2014) 124–135

Table 1
Results of FR and LR in the case of each landslide conditioning factor.

Conditioning factor Class FR Logistic coefficient Conditioning factor Class FR Logistic coefficient

Altitude (m) 0–9.01 0 0.037 Geology 1 0 0.001

9.01–18.02 0 2 0
18.02–45.07 0 3 0
45.07–90.14 0 4 0
90.14–135.21 0 5 0
135.21–198.31 0.19 6 0
198.31–297.46 6.17 7 152.2
297.46–468.73 3.58 8 0
468.73–748.18 0 9 0
748.18–2298.62 0 10 0
Curvature Concave 1.79 0.113 11 0
Flat 0.08 12 0
Convex 1.64 13 0
Distance from river (m) 10 m 0 −0.035 14 0.84
20 m 0 15 0
40 m 0 16 0
100 m 0.62 17 0
200 m 1.01 18 0
400 m 2.16 19 0
1000 m 1.35 20 4.13
2000 m 0.18 21 0
4000 m 0 22 0
8000 m 0 23 5.63
SPI 5.99–7.39 1.34 −0.070 24 0
7.39–8.60 0.19 25 0
8.60–9.30 0 26 0.31
9.30–9.87 0.40 27 0
9.87–10.31 0.91 28 0
10.31–10.76 2.04 29 0
10.76–11.20 1.38 Slope (degree) 0 0.07 0.022
11.20–11.78 2.10 0–1.04 0
11.78–12.54 1.01 1.04–3.12 0.26
12.54–22.21 0.41 3.12–5.21 0.47
Rainfall (mm) 1752–2949 0 0.013 5.21–7.64 1.45
2949–2996 1.34 7.64–10.42 4.30
2996–3199 8.13 10.42–13.90 3.78
3199–3558 0.38 13.90–18.07 2.59
3558–3929 0 18.07–23.98 0.58
3929–4037 0 23.98–88.62 0.30
4037–4121 0 Soil texture Sandy gravel 2.76 0.031
4121–4240 0 Gravel sandy 0
4240–4420 0 Sandy clay 0
4420–4803 0 Clay 0
Soil type Pumiceous 3.185 0.001 Clay sand 0
Aspect −0.005
Flat 0.06
Dasite Phorphyry 68.03 North 1.08
Andesite Northeast 2.93
Miocene Granitic 0 East 6.85
Lithic Crystal Tuff 0 Southeast 1.67
Aluvium 0 South 0.76
Permian Metamorphic 0 Southwest 0.67
Colovium 0 West 1.84
Permian Limestone 0 Northwest 0.91
LULC Primary forest 1.01 −0.016 North 0.73
Secondary forest 1.40 Distance from lineament (m) 30–70 0 0.011
Village 0 70–100 0
Smallholder plantation 0 100–160 0
Open mining 0 160–220 0
Housing 0 220–290 0
Irrigated rice 0 290–380 0
Bush/shrub 0 380–520 0
PGA (8) 1 0 0.018 520–800 0
2 1.67 800–2000 0
3 4.07 2000–4000 4.85
4 0 4000–8000 0.68
8000–16,000 0
16,000–30,000 0
TWI 0–3.74 1.60 0.024
3.74–4.46 1.61
4.46–5.10 2.17
5.10–5.89 2.07
5.89–6.61 1.53
6.61–7.55 0.22
7.55–8.41 0.33
 Z. Umar et al. / Catena 118 (2014) 124–135 127

Table 1 (continued)
Conditioning factor Class FR Logistic coefficient Conditioning factor Class FR Logistic coefficient

TWI 8.41–9.28 0.20

9.28–10.35 0
10.35–18.34 0

approaches such as analytic hierarchy process (AHP) the results will is 87.25%. The average rainfall in 2010 in this area is 427.7 mm
mostly rely on the expert's knowledge and what is appropriate for (Climatological Station, Sicincin, 2011).
regional studies (Subramanian and Ramanathan, 2012). The require-
ment of hazard modeling is to find the global solution, not the method 2.2. Geology
that can be used just for one specific region. Similarly, fuzzy logic is
another method which has more understandable structure than the Rock formations in the southern district of Padang Pariaman consist
ANN. This method has been employed in several landslide applications of alluvium silt, sand and gravel (Leo et al., 1980). Also the rock forma-
and the results depend on the expert's knowledge which leads to a high tion in the Eastern and Northern parts of this area is a very complex
degree of uncertainty (Tilmant et al., 2002). form of volcanic rocks such as hornblende Hypersthene pumiceous
It is one of the important requirements for most multivariate Tuff, Fumiceous Tuff and andesite (Andesite). According to Westerveld
statistical approaches to define strict assumptions prior to the study (1953) deposition tuff was probably derived from the last eruption of
(Benediktsson et al., 1990). Moreover, it is difficult to use them for the caldera Maninjau or fissure in Sumatera. In several locations pumice
real-life applications. The logistic regression (LR) could overcome tuff and andesite are mixed with layers of sand and gravel that are less
these drawbacks and produce an easy way of analysis which doesn't dense. This situation allows the disbursement during the tremor.
require prior assumption and can be combined with other bivariate
statistical analysis (BSA) methods such as frequency ratio (FR) 2.3. Tectonic setting of West Sumatera
(Ayalew and Yamagishi, 2005). The FR is another robust statistical
method which is able to perform BSA (Lee and Dan, 2005). This method The island of Sumatera sits on the top of the Southeast Asian plate,
is capable to assess the impact of the classes of each conditioning factor which overrides the subducting Indian and Australian oceanic plates
due to landslide occurrences (Lee and Pradhan, 2006). On the other that convergence obliquely at about 50 to 60 mm/year (Prawirodirdjo
hand, the LR model is capable to perform multivariate statistical analysis et al., 2000). The Sumatran fault accommodates a significant amount
(MSA), however it doesn't analyze the influence of classes of each of the strike-slip component of the oblique convergence between the
conditioning factor individually on landslides (Yilmaz, 2009). Australian/Indian and Eurasian plates (McCaffrey, 1991; Sieh and
This paper aims to propose an ensemble method of FR and LR in Natawidjaja, 2000). The 1900 km long Sumatran fault zone (SFZ) runs
order to utilize both advantages of these methods and overcome their along the back-bone of Sumatra, within or near the active-volcanic
weak points. So the combination of these two statistical methods can arc. The Sumatran fault zone is highly segmented, and hence consists
generate the complete model that can assess the impact of conditioning of major segments ranging in length from 35 km to 200 km. Four of
factors and also the impact of classes of each conditioning factor on them are located in the province of West Sumatera, namely Fault
landslide occurrence. As mentioned earlier, Padang is highly susceptible Sumpur along 35 km, Sianok along 90 km, Sumani along 60 km, and
to landslide occurrence due to severe earthquake, hence, landslide Suliti along 95 km. The Sumatran fault zone poses major hazards, partic-
susceptibility analysis is needed in order to recognize prone areas and ularly to people on and near the active faults. Since the 1890s, about 21
prevent more development and urbanization in those areas. Through major earthquakes have ruptured the segments of the Sumatran fault
proper strategies and plans, many lives can be saved and the degree of zone with magnitudes ranging from 6.5 to 7.7 (Sieh and Natawidjaja,
damages may be decreased as well. The objective of this paper is to 2000). Based on the focal depth and dip angle, subduction zones were
generate a landslide susceptibility map of the Padang Pariaman District divided into interpolate (Megathrust) zones and interplate (Benioff)
using the ensemble FR and LR methods as a simple and understandable zones. Megathrust zone earthquake events occurred at depth less shal-
probabilistic model. low than 50 km, and depth of 50–200 km was represented as Benioff
zone.
2. Study area
3. Data used
2.1. Geographic
3.1. Landslide inventory map
Geographically Padang Pariaman is located between latitude 0° 11′–
0° 49′ South and longitude 98° 36′–100° 28′East, encompassing Landslide inventory maps are considered to be the basis of most of
1328.79 km2 (Fig. 1). In total, the land area of Padang Pariaman is equiv- the susceptibility mapping techniques. In addition, the inventory maps
alent to 3.15% of the land area of the West Sumatra Province. The capital can be used for evaluating and reducing landslide hazard or risk on a
city of Padang Pariaman is Pariaman City which is approximately 55 km regional scale (Wieczorek, 1984). The earthquake which happened on
from Padang, the capital city of West Sumatra. The topography of Wednesday 30th of September 2009 at 17:16 pm with a magnitude of
Padang consists of land located on mainland Sumatra and six small 8 RS in Padang Pariaman has triggered landslides in the area. Most of
islands. Forty percent of the region is located near the beach, with an these landslides occurred in unstable regions that had steep slope
average width of approximately 1–3 km from the beach and an eleva- which were composed by rock weathering. Also it occurred in the area
tion of 0–10 m above the sea level. The remaining 60% of this region is which was made of less compact and less dense soil, such as the region
located in the East, especially in undulating hills with an elevation of composed of Hornblenda Hipersten Pumiceous Tuff unit (Qhpt),
10–1000 m above the sea level. Pumiceous Tuff (Qpt), Andesite and Basalt. Most of these landslides
Padang has a tropical climate with a short dry season and this area is are debris flows.
also affected by the sea breeze. The temperature ranges between 24.4 °C The existing landslide inventory map is very essential for studying
and 25.78 °C. The maximum temperature in general is 31.08 °C while the relationship between the landslide distribution and the condition-
the minimum is 21.34 °C. The highest and lowest temperatures ing factors. To produce a detailed and reliable inventory map, extensive
occur in May and September and the relative humidity for this area field surveys and observations were performed in the study area. A total
128 Z. Umar et al. / Catena 118 (2014) 124–135
 Z. Umar et al. / Catena 118 (2014) 124–135 129

Fig. 2. List of all the conditioning factors (a) DEM, (b) curvature, (c) river, (d) SPI, (e) rainfall, (f) soil type, (g) soil texture, (h) LULC, (i) PGA 8, (j) geology, (k) slope, (l) aspect, (m) distance
from lineament and (n) TWI.

of 89 landslides were identified in the study area by evaluating be used in landslide susceptibility mapping (Liu and De Smedt, 2005),
Quickbird imagery (0.61 m spatial resolution) which has been taken which can be in nominal, ordinal, or scale format (Park et al., 2013).
on April 1, 2011 and by multiple field surveys (Fig. 1). The average For current research the conditioning factors are: altitude, curvature,
area of the landslide was approximately 0.25 ha. For this analysis, the river, stream power index (SPI), rainfall, soil type, soil texture, land
landslide inventory map was randomly divided into two groups 70% use/cover (LULC), peak ground acceleration (PGA) 8, geology, slope,
and 30% of the landslide locations for the purpose of training and testing aspect, lineament, and topographic wetness index (TWI). The list of
respectively. Consequently, the dependent layer was prepared using the the conditioning factors and their characteristics is shown in Table 1.
training dataset. This layer was made up of 0 and 1 values. The value of 1 Each factor was resized to 10 × 10 m grid and the grid of the Padang
represents the existence and the value of 0 illustrates the absence of Pariaman area was constructed by 2795 columns and 2850 rows
landslide over the area. The rest of the landslide events were used for (3,491,989 pixels; 1396 km2). For FR analysis, quantile method was
the purpose of testing. After dividing the landslide locations, both layers used to classify each conditioning factor and the list of all the data layers
were prepared and converted to raster. is illustrated in Fig. 2. Elevation is one of the topographic factors that
affect slope instability (Wan et al., 2012). Other parameters such as
3.2. Landslide conditioning factors precipitation and weathering variations are considered as indirect
factors. In this study, the elevation map was obtained from the altitude
Generally speaking, the conditioning factors are responsible for the which was generated using survey of Indonesia topographic map sheets
occurrence of landslide in specific region. These factors are needed to of the Padang Pariaman area (1984, 1985 and 1996). This layer was
130 Z. Umar et al. / Catena 118 (2014) 124–135

Fig. 2 (continued).

classified into 10 classes (Fig. 2a). In the curvature map, three classes (Chen and Liu, 2012). Eqs. (1) and (2) represent the mathematical
were considered; convex, flat and concave (Fig. 2b). formulation of IDW:
In the case of the distance from the river map, ten buffer classes
(10 m, 20 m, 40 m, 100 m, 200 m, 400 m, 1000 m, 2000 m, 4000 m X
N
^ ¼
R wi Ri ð1Þ
and 8000 m) were made using the buffer tool in ArcGIS 10 software p
i¼1
(Fig. 2c). In the SPI (5.99–22.21) and rainfall (1752–4803) maps, ten
categories were created for each of the analysis (Fig. 2d and e). In this
research, the average rainfall data were collected from the meteorolog-
d−∝
ical stations surrounding the study area (Pourghasemi et al., 2012; Wan wi ¼ XNi −∝ ð2Þ
et al., 2010a). There are various geostatistical interpolation methods d
i¼1 i
that exist to calculate rainfall such as the average arithmetic, Thiessen
polygon method, Isohyetal method, inverse distance weight (IDW), where R ^ means the unknown rainfall data (mm); Ri means the rainfall
p
and Kriging. In the current research, IDW was applied using line density data of known rainfall stations (mm); N means the amount of rainfall sta-
operations in ArcGIS 10. The general assumption of IDW is that the value tions; wi means the weighting of each rainfall station; di means the dis-
of an unsampled point is the weighted average of known values within tance from each rainfall station to the unknown site; and ∝ means the
the neighborhood. Therefore, the values from a scattered set of known power, and is also a control parameter. Using the mentioned equations,
points can be utilized to recognize and assign the values to unknown the rainfall average data for the whole study area was interpolated.
points. It can be used to compute the unknown spatial rainfall data Nine categories of soil types were recognized: Pumiceous, Dasite
from the known data of sites that are adjacent to the unknown site Phorphyry, Andesite, Miocene Granitic, Lithic Crystal Tuff, Aluvium,
 Z. Umar et al. / Catena 118 (2014) 124–135 131

Permian Metamorphic, Colovium, and Permian Limestone (Fig. 2f). The was applied in order to evaluate the landslide susceptibility of the
soil texture is related to the clay fractions, silt and sand. The finer the area based on the observed spatial relationship between landslide loca-
grain of soil surface texture, the more ability to absorb water and there- tions and conditioning factors. In general, the landslides that will occur
fore more landslide occurrence is expected. In the current research, the in the future have the same condition as past landslides (Lee and
soil texture is obtained from the digitized drainage layer geomorpholog- Pradhan, 2006). The FR is the ratio of the probability of the presence
ical map of Padang sheet Sumatera. There are five classes of soil texture to the absence of landslide occurrence (Lee and Pradhan, 2007;
that exist in the study area: sandy gravel, gravel sandy, sandy clay, clay, Yilmaz, 2009). In the current research, four PGA (7.5, 8, 8.6 and 9)
and clay sand. The distribution of the soil texture is illustrated in Fig. 2g. maps were acquired for different years, and subsequently a PGA map
Generally, the landuse/landcover (LULC) analysis reflects the relation- of 8 for the 2009 earthquake was prepared. Therefore, PGA 8 was used
ships between land use and landslides. In this study, topographic map to run the model by forming the structure of the final probability equa-
and Landsat 7 ETM+ satellite image were acquired on Mar 1, 2009 in tion. Similarly, for the other three PGA the same equation was used to
order to map the LULC. In the LULC map, eight classes were found: pri- run the model by changing the different PGA values. The purpose was
mary forest, secondary forest, village, small holder plantation, open to generate the fixed model that can be used in various events and pro-
mining, housing, irrigated rice, and bush (shrub) (Fig. 2h). duce rapid, accurate and trustable results. The spatial relationship be-
The peak ground acceleration (PGA) map of Padang Pariaman tween the conditioning factor and landslides was measured through
District was constructed through the following steps: the FR model. The higher the FR ratio is, the stronger the relationship be-
tween the landslide occurrence and the conditioning factor (Pradhan
a. Firstly, the district of Padang Pariaman was divided by a grid and Lee, 2010b).
size of 3.5 km × 3.5 km. Once the area was divided by a grid The classified conditioning factors are illustrated in Fig. 2. Then, the
size 3.5 × 3.5 km, the center of each pixel was determined. FR was performed and the range of FR was normalized. The normalized
b. Then, the calculation of PGA was done using a global attenuation FR values were used to reclassify all the conditioning factors to use them
adapted from Fukushima and Tanaka (1990): in the LR analysis. The normalization of all independent variables is nec-
essary which facilitate the final analysis and interpretation (Ayalew and
Yamagishi, 2005). For normalization process, a common method was
h i used for this purpose (Choi et al., 2010). For example, for yi(i = 1, 2,.
0:41MS
Log10 A ¼ 1:30 þ 0:41MS − log R þ 0:032  10 −0:0034R ð3Þ
… n):

where;
yi −ymin
yi ¼ ð4Þ
ymax −ymin
A is the mean peak acceleration (in cm/s2)
MS is the magnitude of the earthquake
R is the shortest distance between the site and fault rupture where, yi means the normalized values of yi, and ymin and ymax represent
(km) or the shortest distance from the location to the earth- the minimum and maximum values of yi respectively.
quake source (km).
4.2. Application of logistic regression (LR)
In this research, four PGA layers from different years have been
The LR is considered as one of the popular MSA methods which
collected. The PGA 8 was measured from the earthquake that took place
permits one to make a multivariate regression relationship between a
in 2009 in Padang and the model has been generated for this PGA. Subse-
dependent variable and numerous conditioning factors (Pradhan and
quently, four classes were defined for PGA 8 as can be seen in Fig. 2i.
Lee, 2010b). McFadden (1974) introduced LR for the first time which
Twenty nine geology classes have been found in the study area as
measured the probability of any disaster in a specific area using specific
can be seen in Fig. 2j. Based on this map and the observations in the
formula and the conditioning factors. The dependent data is made up of
field, the study area was mostly formed by various volcanic, sedimenta-
0 and 1 values which show the absence and presence of landslides
ry, metamorphic and igneous rocks. The various classes of the lithology
respectively. All the datasets contain dependent layer and all the condi-
are illustrated in Fig. 2k. One of the most important conditioning factors
tioning factors were converted to ASCII format to be used in SPSS
in the slope stability analysis is the slope angle. For this reason, the slope
program. The LR was applied using SPSS V.19 software and the coeffi-
map of the study area was prepared from the altitude, and it was
cients were measured. The list of all LR coefficients can be seen in
divided into 10 slope categories using quantile method (Pourghasemi
Table 1. The higher the LR coefficient, the higher impacts on landslide
et al., 2012) (Fig. 2k). In landslide susceptibility studies, slope aspect is
occurrence can be expected. The derived LR coefficients were used in
considered to be an important conditioning factor (Baeza and
order to calculate the probability (p) of landslide occurrence using Eq.
Corominas, 2001). The reason is, aspect-related parameters such as
(5):
exposure to sunlight and drying winds, control the concentration of
the soil moisture which is directly related to the landslide occurrence  −z 
(Magliulo et al., 2008). In the case of the aspect map, ten classes have p ¼ 1= 1 þ e ð5Þ
been made (flat, North, Northeast, East, Southeast, South, Southwest,
West, Northwest and North) (Fig. 2l). The lineament density map was where, p is the probability of landslide occurrence. In this situation,
obtained by using the digitized drainage layer made by the Survey of the value p is the estimated probability of landslide occurrence. The
Indonesia (1984, 1985 and 1996) and has been divided into twelve clas- probability varies from 0 to 1 on an S-shaped curve and z is the linear
ses. The distribution of the lineament density is illustrated in Fig. 2m. In combination. It follows that logistic regression involves fitting an
the TWI (0–18.34) map, ten categories were created (Fig. 2d and e). equation of the following form to the data:

4. Methodology
z ¼ b þ b1 x1 þ b2 x2 þ b3 x3 þ bn xn ð6Þ
4.1. Application of frequency ratio (FR)
where b° is the intercept of the model, bi (i = 0, 1, 2, …, n) represents
The LSM can be done by several methods such as statistical, deter- the coefficients of the LR model, and xi (i = 0, 1, 2, …, n) denotes the
ministic and heuristic methods. In the current research, the FR model conditioning factors (Lee and Sambath, 2006).
132 Z. Umar et al. / Catena 118 (2014) 124–135

5. Results The acquired landslide probability map showed that the probability
of landslide occurrence is large in high slope and high elevated areas. In
The FR method was produced using the weights for each class of order to perform LSM, the probability map should be divided into differ-
each conditioning factor. Through analyzing the relationship between ent categories. In the literature, it is possible to see different types of
13 conditioning factors and landslide occurrence, the FR ratio was classification schemes such as standard deviation, quantile, natural
calculated (Table 1). break and equal interval. In the current study the best results are
Table 1 represents the relationship between landslide event and the achieved through the quantile method. The outputs of other classifica-
classes of each conditioning factor. It can be seen that in the case of the tion methods showed the susceptible areas with high degree of exag-
relationship between landslide occurrence and altitude, landslide most- geration and larger part of the study area fell into the high susceptible
ly occurred in the elevation range of 198.31–297.46 m. It showed that zone. Finally, a landslide susceptibility map of PGA 8 was obtained and
the probability of landslide occurrence is very low in low elevation the study area was divided into five classes of landslide susceptibility:
areas. In the case of the curvature, the ratio was N1 for the class of very low (0–0.29), low (0.29–0.37), medium (0.37–0.45), high (0.45–
concave, and the other classes of flat and convex have ratios of 0.08 0.52) and very high (0.52–1). The derived landslide susceptibility map
and 1.64 respectively. The most effective class of distance from river is shown in Fig. 4.
was 400 m with a ratio of 2.16. For SPI, the FR ratio was highest (2.10) The other three equations were applied by changing the PGA (7.2,
for the class of 11.20–11.78, and it was lowest (0) for the class of 8.6 and 9) and landslide susceptibility map of each can be seen in
8.60–9.30. The highest FR ratio for rainfall classes as main contributors Fig. 4. All the four susceptibility maps represent almost the same condi-
of landslide belonged to the area of 2996–3199 mm precipitation with tion for the study area. Based on the acquired landslide susceptibility
a ratio of 8.13. In the case of soil types, the highest FR ratio (68.03) map, most parts of the study area are located in very low susceptible
was related to the Dasite Phorphyry which showed the significant zones. The high and very high landslide susceptible areas acquired
influence of this soil type on the landslide occurrence. All the soil type from PGA 8 covered 19.4% of the area and mostly it is located in the
classes have zero FR value, expect for the class of Pumiceous. North and North-East parts of Padang. The area which contained clay
The LULC map is one of the conditioning factors that have direct soil type, very high elevation and precipitation was classified as the
impact on the occurrence of the landslides. For LULC, the highest FR most susceptible zone. Also results revealed that there is a negative
ratios of 1.40 and 1.01 were achieved for secondary forest and primary correlation between landslide occurrence and river, SPI, LULC and
forest classes which amplified the landslide occurrences. In terms of aspect. The proposed ensemble method is recommended for landslide
PGA, the third class of PGA 8 produced the highest FR ratio of 4.07. In studies, as changing the whole dataset and running new equation are
the case of the geology, all the classes have a ratio of zero, except the not required. The acquired PGA can be entered in the pre-developed
classes of 7, 14, 20, 23 and 26. For the areas with the slope angles model and in a very short time the results can be generated.
7.64°–10.42°, the ratio was 4.30, which illustrated that the higher The proposed model has been evaluated using area under the (ROC)
slope angle increases the susceptibility of the area to the landslide curve. Both success and prediction rates were measured to examine
occurrence. The most effective class of the soil texture was sandy gravel, the reliability and efficiency of the landslide probability map. The
with the FR ratio of 2.76 and other three classes acquired zero as FR validation has been done by comparing the existing landslide data
ratio. This indicates that the most influential class was sandy gravel with the produced landslide probability map (Lee and Pradhan, 2007;
while the others are not effective in landslide occurrence. Also the Pourghasemi et al., 2012). The training and testing datasets were used
impact of the aspect was evaluated as contributor in landslide occur- to acquire success and prediction rates respectively. The training dataset
rence. The ratio was highest (6.85) for the class of East, and it was lowest (70% of the inventory landslide locations i.e. 61 landslide locations) was
(0.06) for the class of flat. For the lineament, the FR ratio was zero for all used to get the success rate. The success rate was 84% and the curve can
the classes, except the classes 2000–4000 and 4000–8000 which have be seen in Fig. 5. While the training dataset was used to assess the
ratios of 4.85 and 0.68 respectively. In the case of TWI, the FR ratio model, it cannot present the real efficiency of the developed model.
was highest (2.17) for the class of 4.46–5.10, and it was lowest (0) for The prediction rate shows how well the model can predict the landslide
the classes of 9.28–10.35 and 10.35–18.34. occurrence in an area (Tien Bui et al., 2012a). The testing dataset
The MSA was performed using LR and the relationship between the
landslide occurrence and landslide conditioning factors was assessed in
SPSS V.19 software. The LR coefficients are listed in Table 1. As can be
seen in Table 1, curvature showed the high and positive correlation
with landslide occurrence as it could acquire the highest LR coefficient
of 0.113. Similarly, altitude and soil texture are the other effective
conditioning factors, by the LR coefficient of 0.037 and 0.031 respectively.
In order to obtain the landslide probability index the regression
coefficients for each conditioning factor were entered in Eq. (7):

Z ¼ −22:742 þ ð0:113 curvatureÞ−ð0:005 aspectÞ

þð0:011 distance from lineamentÞ−ð0:035 distance from riverÞ
þð0:037 altitudeÞ þ ð0:001 geologyÞ−ð0:016 LULCÞ
ð7Þ
þð0:018 PGA8Þ þ ð0:013 rainfallÞ þ ð0:022 slopeÞ

þð0:031 soil textureÞ þ ð0:001 soil typeÞ

−ð0:07 SPIÞ þ ð0:024 TWIÞ:

Finally, the probability index was calculated which ranges from 0 to

0.99. The landslide probability map for PGA 8 can be seen in Fig. 3 where
the index illustrates the predicted probabilities of landslide for each
pixel in the presence of a given set of conditioning factors. Fig. 3. The probability map obtained from ensemble FR and LR methods.
 Z. Umar et al. / Catena 118 (2014) 124–135 133

Fig. 4. Landslide susceptibility map produced from ensemble FR and LR methods for (a) PGA 7.2, (b) 8, (c) PGA 8.6 and (d) PGA 9.

(selected by 30% of landslide areas of total number 26 landslide sam- which shows that the model was performed perfectly. So the closer the
ples) which was not used in the training process was used to measure value of AUC to the 1.0 is, the model is considered to be more precise
the prediction rate. So the model's generalization ability can be assessed and trustable (Tien Bui et al., 2012b). The prediction rate was 78% and
through this method (Maier and Dandy, 2000). The range of the AUC is shown in Fig. 5.
varies from 0.5 to 1.0, as the value of 1.0 represents the highest accuracy

Fig. 5. ROC curve. Fig. 6. Prediction rate for models derived from PGA 7.2, 8.6 and 9.
134 Z. Umar et al. / Catena 118 (2014) 124–135

Validation also has been done for the other three models that were Benediktsson, J., Swain, P.H., Ersoy, O.K., 1990. Neural network approaches versus
statistical methods in classification of multisource remote sensing data. IEEE Trans.
acquired from PGA 7.2, 8.6 and 9. The prediction rate was measured to Geosci. Remote Sens. 28, 540–552.
examine the capability of the generated equation. The prediction rates Bui, D.T., Pradhan, B., Lofman, O., Revhaug, I., Dick, Ø.B., 2013. Regional prediction of land-
for the models made by PGA 7.2, 8.6 and 9 were 79%, 78% and 81% slide hazard using probability analysis of intense rainfall in the Hoa Binh province.
Vietnam. Nat. Hazards 66, 707–730.
respectively (Fig. 6). It can be seen that the proposed ensemble Chau, K., Wu, C., Li, Y., 2005. Comparison of several flood forecasting models in Yangtze
method was able to predict landslides due to different values of River. J. Hydrol. Eng. 10, 485–491.
PGA. The generated equation can produce the results and detect the Chen, F.W., Liu, C.W., 2012. Estimation of the spatial rainfall distribution using inverse
distance weighting (IDW) in the middle of Taiwan. Paddy Water Environ. 10, 209–222.
susceptible areas in a very short time.
Chi, K.H., Park, N.W., Chung, C.J., 2002. Fuzzy logic integration for landslide hazard
mapping using spatial data from Boeun, Korea. International Archives of Photo-
6. Conclusions grammetry Remote Sensing and Spatial. Inf. Sci. 34, 54–59.
Choi, J., Oh, H.J., Won, J.S., Lee, S., 2010. Validation of an artificial neural network model for
landslide susceptibility mapping. Environ. Earth Sci. 60, 473–483.
Earthquake is one of the main contributors in landslide occurrence Fukushima, Y., Tanaka, T., 1990. A new attenuation relation for peak horizontal acceleration
which results in severe human casualties, property losses and environ- of strong earthquake ground motion in Japan. Bull. Seismol. Soc. Am. 80, 757–783.
Ghalkhani, H., Golian, S., Saghafian, B., Farokhnia, A., Shamseldin, A., 2012. Application of
mental degradation. Earthquake induced landslides can cause extensive surrogate artificial intelligent models for real‐time flood routing. Water Environ. J.
and significant damages to both lives and properties. The geographical http://dx.doi.org/10.1111/j.1747-6593.2012.00344.x.
condition of Padang Pariaman, Indonesia is formed by mountains that Gupta, R.P., Kanungo, D., Arora, M.K., Sarkar, S., 2008. Approaches for comparative evaluation
of raster GIS-based landslide susceptibility zonation maps. Int. J. Appl. Earth Obs. Geoinf.
easily fall down especially after earthquake and heavy rain. An 8 RS 10, 330–341.
earthquake struck West Sumatra and its surrounding areas on Wednes- Islam, I., Khan, H., 1986. Spatial patterns of inequality and poverty in Indonesia. Bull.
day, 30 September 2009. Due to this earthquake a huge number of Indones. Econ. Stud. 22, 80–102.
Jebur, M.N., Pradhan, B., Tehrany, M.S., 2013a. Detection of vertical slope movement in
landslides happened in the North-East of this area which made the
highly vegetated tropical area of Gunung pass landslide, Malaysia, using L-band
local governments investigate an appropriate and rapid solution to InSAR technique. Geosci. J. 1–8. http://dx.doi.org/10.1007/s12303-013-0053-8.
detect the landslide susceptible areas. Over many years, geographic Jebur, M.N., Pradhan, B., Tehrany, M.S., 2013b. Using ALOS PALSAR derived high-
resolution DInSAR to detect slow-moving landslides in tropical forest: Cameron
information system (GIS) and remote sensing (RS) have become
Highlands, Malaysia. Geomat. Nat. Haz. Risk 1–19. http://dx.doi.org/10.1080/
integral tools for the evaluation of natural hazard phenomena. In the 19475705.2013.860407.
current research an ensemble method of FR and LR was applied for Karig, D.E., 1974. Evolution of arc systems in the western Pacific. Annu. Rev. Earth Planet.
LSM. Four PGA were collected and the one which was related to the Sci. 2, 51.
Konadu, D., Fosu, C., 2009. Digital elevation models and GIS for watershed model-
2009 earthquake was used to generate the model equation. The aim ling and flood prediction—a case study of Accra Ghana. Appropriate Technolo-
was to establish a pre-developed model for future studies which can gies for Environmental Protection in the Developing World. Springer,
produce the LSM just by entering the PGA to the formula. The FR was Netherlands, pp. 325–332.
Lee, S., Dan, N.T., 2005. Probabilistic landslide susceptibility mapping in the Lai Chau prov-
chosen to perform BSA and extract the spatial correlation between ince of Vietnam: focus on the relationship between tectonic fractures and landslides.
classes of each conditioning factor and landslide occurrence. The output Environ. Geol. 48, 778–787.
of the FR was reclassified based on the weights of each class and was Lee, S., Pradhan, B., 2006. Probabilistic landslide hazards and risk mapping on Penang
Island, Malaysia. J. Earth Syst. Sci. 115, 661–672.
used as an input for LR. Subsequently, the LR model analyzed the impact Lee, S., Pradhan, B., 2007. Landslide hazard mapping at Selangor, Malaysia using frequen-
of each conditioning factor on the landslides. The probability map was cy ratio and logistic regression models. Landslides 4, 33–41.
extracted and then reclassified into five landslide susceptibility zones Lee, S., Sambath, T., 2006. Landslide susceptibility mapping in the Damrei Romel area,
Cambodia using frequency ratio and logistic regression models. Environ. Geol. 50,
using quantile method. The validation results showed that the ensemble
847–855.
model could provide 84% and 78% success and prediction rates respec- Leo, G.W., Hedge, C.E., Marvin, R.F., 1980. Geochemistry, strontium isotope data, and
tively. Once the model is generated, other three PGA (7.2, 8.6 and 9) potassium–argon ages of the andesite–rhyolite association in the Padang area, west
Sumatra. J. Volcanol. Geotherm. Res. 7, 139–156.
were used to model the other three LSM. The prediction rates for the
Liu, Y., De Smedt, F., 2005. Flood modeling for complex terrain using GIS and remote
models made by PGA 7.2, 8.6 and 9 were 79%, 78% and 81% respectively. sensed information. Water Resour. Manag. 19, 605–624.
The results proved the efficiency of the proposed method and all the Luzi, L., Pergalani, F., Terlien, M.T.J., 2000. Slope vulnerability to earthquakes at sub-
maps visually represented almost the same susceptible areas which regional scale, using frequency techniques and geographic information systems.
Eng. Geol. 58, 313–336.
were located in the North and North-East of Padang. Through the LSM Magliulo, P., Di Lisio, A., Russo, F., Zelano, A., 2008. Geomorphology and landslide suscepti-
the most effective landslide conditioning factors were detected e.g. bility assessment using GIS and bivariate statistics: a case study in southern Italy. Nat.
curvature, altitude and soil texture which had positive and high correla- Hazards 47, 411–435.
Maier, H.R., Dandy, G.C., 2000. Neural networks for the prediction and forecasting of
tion with landslide occurrence. Through the analysis that has been done water resources variables: a review of modelling issues and applications. Environ.
in the current research, the impact and the relation of each conditioning Modell. Softw. 15, 101–124.
factor on landslide have been assessed. The current research provides a McCaffrey, R., 1991. Slip vectors and stretching of the Sumatran fore arc. Geology 19,
881–884.
valuable characterization of triggered landslides in Padang, Indonesia McFadden, D., 1974. Conditional logit analysis of qualitative choice analysis. In: Zarembka,
that can be used to predict the characteristics of landslides triggered P. (Ed.), Frontiers in Econometrics. Academic Press, New York, pp. 105–142.
by future earthquakes in that area. Mohammady, M., Pourghasemi, H.R., Pradhan, B., 2012. Landslide susceptibility mapping
at Golestan Province, Iran: a comparison between frequency ratio, Dempster–Shafer,
and weights-of-evidence models. J. Asian Earth Sci. 61, 221–236.
References Oh, H.J., Pradhan, B., 2011. Application of a neuro-fuzzy model to landslide-
susceptibility mapping for shallow landslides in a tropical hilly area. Comput.
Akgun, A., Sezer, E.A., Nefeslioglu, H.A., Gokceoglu, C., Pradhan, B., 2012. An easy-to-use Geosci. 37, 1264–1276.
MATLAB program (MamLand) for the assessment of landslide susceptibility using a Oh, H.J., Lee, S., Soedradjat, G.M., 2010. Quantitative landslide susceptibility mapping at
Mamdani fuzzy algorithm. Comput. Geosci. 38, 23–34. Pemalang area, Indonesia. Environ. Earth Sci. 60, 1317–1328.
Ayalew, L., Yamagishi, H., 2005. The application of GIS-based logistic regression for Ozdemir, A., Altural, T., 2012. A comparative study of frequency ratio, weights of evidence
landslide susceptibility mapping in the Kakuda–Yahiko Mountains, Central Japan. and logistic regression methods for landslide susceptibility mapping: Sultan Moun-
Geomorphology 65, 15–31. tains, SW Turkey. J. Asian Earth Sci. 64, 180–197.
Baeza, C., Corominas, J., 2001. Assessment of shallow landslide susceptibility by means of Park, S., Choi, C., Kim, B., Kim, J., 2013. Landslide susceptibility mapping using frequency
multivariate statistical techniques. Earth Surf. Process. Landf. 26, 1251–1263. ratio, analytic hierarchy process, logistic regression, and artificial neural network
Bai, S., Wang, J., Bell, R., Glade, T., 2012. Distribution and susceptibility assessments methods at the Inje area, Korea. J. Asian Earth Sci. 68, 1443–1464.
of landslide triggered by Wenchuan earthquake at Longnan. Proceedings of the Pourghasemi, H.R., Pradhan, B., Gokceoglu, C., 2012. Application of fuzzy logic and analyt-
International Conference on Informatics, Cybernetics, and Computer Engineering ical hierarchy process (AHP) to landslide susceptibility mapping at Haraz watershed,
(ICCE2011), Melbourne, Australia. Iran. Nat. Hazards 63, 965–996.
Barredo, J., Benavides, A., Hervás, J., van Westen, C.J., 2000. Comparing heuristic landslide Pourghasemi, H.R., Jirandeh, A.G., Pradhan, B., Xu, C., Gokceoglu, C., 2013. Landslide sus-
hazard assessment techniques using GIS in the Tirajana basin, Gran Canaria Island, ceptibility mapping using support vector machine and GIS at the Golestan Province,
Spain. Int. J. Appl. Earth Obs. Geoinf. 2, 9–23. Iran. J. Asian Earth Sci. 122, 349–369.
 Z. Umar et al. / Catena 118 (2014) 124–135 135

Pradhan, B., 2010. Remote sensing and GIS-based landslide hazard analysis and cross- Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., Dick, O.B., 2012. Spatial prediction of
validation using multivariate logistic regression model on three test areas in landslide hazards in Hoa Binh province (Vietnam): a comparative assessment of
Malaysia. Adv. Space Res. 45, 1244–1256. the efficacy of evidential belief functions.a comparative assessment of the efficacy
Pradhan, B., 2011. Use of GIS-based fuzzy logic relations and its cross application to of evidential belief functions and fuzzy logic models. Catena 96, 28–40.
produce landslide susceptibility maps in three test areas in Malaysia. Environ. Earth Tilmant, A., Vanclooster, M., Duckstein, L., Persoons, E., 2002. Comparison of fuzzy and
Sci. 63, 329–349. nonfuzzy optimal reservoir operating policies. J. Water Resour. Plan. Manag. 128,
Pradhan, B., 2013. A comparative study on the predictive ability of the decision tree, 390–398.
support vector machine and neuro-fuzzy models in landslide susceptibility mapping Van Asch, T.W., Buma, J., Van Beek, L., 1999. A view on some hydrological triggering
using GIS. Comput. Geosci. 51, 350–365. systems in landslides. Geomorphology 30, 25–32.
Pradhan, B., Buchroithner, M.F., 2010. Comparison and validation of landslide susceptibil- Van Westen, C.J., 2000. The modelling of landslide hazards using GIS. Surv. Geophys. 21,
ity maps using an artificial neural network model for three test areas in Malaysia. 241–255.
Environ. Eng. Geosci. 16, 107–126. Van Westen, C., Rengers, N., Soeters, R., 2003. Use of geomorphological information in
Pradhan, B., Lee, S., 2010a. Delineation of landslide hazard areas on Penang Island, indirect landslide susceptibility assessment. Nat. Hazards 30, 399–419.
Malaysia, by using frequency ratio, logistic regression, and artificial neural network van Westen, C.J., Castellanos, E., Kuriakose, S.L., 2008. Spatial data for landslide suscepti-
models. Environ. Earth Sci. 60, 1037–1054. bility, hazard, and vulnerability assessment: an overview. Eng. Geol. 102, 112–131.
Pradhan, B., Lee, S., 2010b. Regional landslide susceptibility analysis using back- Wan, S., Lei, T.C., 2009. A knowledge-based decision support system to analyze the debris-
propagation neural network model at Cameron Highland, Malaysia. Landslides 7, flow problems at Chen-Yu-Lan River, Taiwan. Knowl.-Based Syst. 22, 580–588.
13–30. Wan, S., Lei, T., Chou, T.Y., 2010a. A novel data mining technique of analysis and classifi-
Pradhan, B., Youssef, A.M., Varathrajoo, R., 2010c. Approaches for delineating landslide cation for landslide problems. Nat. Hazards 52, 211–230.
hazard areas using different training sites in an advanced artificial neural network Wan, S., Lei, T.C., Chou, T.Y., 2010b. An enhanced supervised spatial decision support
model. Geospatial. Inf. Sci. 13 (2), 93–102. system of image classification: consideration on the ancillary information of paddy
Pradhan, B., Mansor, S., Pirasteh, S., Buchroithner, M.F., 2011. Landslide hazard and risk rice area. Int. J. Geogr. Inf. Sci. 24, 623–642.
analyses at a landslide prone catchment area using statistical based geospatial Wan, S., Lei, T.C., Chou, T.Y., 2012. A landslide expert system: image classification through
model. Int. J. Remote. Sens. 32 (14), 4075–4087. integration of data mining approaches for multi-category analysis. Int. J. Geogr. Inf.
Prawirodirdjo, L., Bock, Y., Genrich, J., Puntodewo, S., Rais, J., Subarya, C., Sutisna, S., 2000. Sci. 26, 747–770.
One century of tectonic deformation along the Sumatran fault from triangulation and Westen, C.V., Terlien, M., 1996. An approach towards deterministic landslide hazard
Global Positioning System surveys. J. Geophys. Res. Solid Earth 105, 28343–28361. analysis in GIS. A case study from Manizales (Colombia). Earth Surf. Process. Landf.
Saha, A.K., Gupta, R.P., Sarkar, I., Arora, M.K., Csaplovics, E., 2005. An approach for GIS- 21, 853–868.
based statistical landslide susceptibility zonation—with a case study in the Himalayas. Westerveld, J., 1953. Eruptions of acid pumice tuffs and related phenomena along the
Landslides 2, 61–69. Great Sumatran fault-trough system. Proceedings of the Pacific Science Congress,
Sarkar, S., Kanungo, D., 2004. An integrated approach for landslide susceptibility mapping New Zealand.
using remote sensing and GIS. Photogramm. Eng. Remote. Sens. 70, 617–626. Wieczorek, G., 1984. Preparing a detailed landslide-inventory map for hazard evaluation
Sidle, R.C., Ochiai, H., 2006. Landslides: Processes, Prediction, and Land Use. American and reduction. Bull. Assoc. Eng. Geol. 21, 337–342.
Geophysical Union, Washington (312 pp.). Wilkinson, S., Alarcon, J., Mulyani, R., Whittle, J., Chian, S., 2012. Observations of damage
Sieh, K., Natawidjaja, D., 2000. Neotectonics of the Sumatran fault, Indonesia. J. Geophys. to buildings from M w 7.6 Padang earthquake of 30 September 2009. Nat. Hazards
Res. Solid Earth 105, 28295–28326. 63, 521–547.
Subramanian, N., Ramanathan, R., 2012. A review of applications of analytic hierarchy Yilmaz, I., 2009. Landslide susceptibility mapping using frequency ratio, logistic regres-
process in operations management. Int. J. Prod. Econ. 138, 215–241. sion, artificial neural networks and their comparison: a case study from Kat landslides
Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., 2012a. Landslide susceptibility assess- (Tokat—Turkey). Comput. Geosci. 35, 1125–1138.
ment in Vietnam using support vector machines, decision tree, and Naive Bayes Yin, Y., Wang, F., Sun, P., 2009. Landslide hazards triggered by the 2008 Wenchuan
Models. Math. Probl. Eng. 2012, 1–26. earthquake, Sichuan, China. Landslides 6, 139–152.
Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., Dick, O.B., 2012b. Landslide susceptibility Yin, J., Chen, J., Xu, X., Wang, X., Zheng, Y., 2010. The characteristics of the landslides
mapping at Hoa Binh province (Vietnam) using an adaptive neuro-fuzzy inference triggered by the Wenchuan Ms N 8.0 earthquake from Anxian to Beichuan. J. Asian
system and GIS. Comput. Geosci. 45, 199–211. Earth Sci. 37, 452–459.
Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., Dick, O.B., 2012c. Landslide susceptibility Zhou, J.W., Cui, P., Yang, X.G., 2013. Dynamic process analysis for the initiation and
assessment in the Hoa Binh province of Vietnam: a comparisonof the Levenberg- movement of the Donghekou landslide-debris flow triggered by the Wenchuan
Marquardt and Bayesian regularized neural networks. Geomorphology 171, 12–29. earthquake. J. Mt. Sci. 76, 70–84.