Producing a landslide inventory map using pixel-based and object-oriented
approaches optimized by Taguchi method

Vahid Moosavi, Ali Talebi, Bagher Shirmohammadi

PII: S0169-555X(13)00470-4
DOI: doi: 10.1016/j.geomorph.2013.09.012
Reference: GEOMOR 4492

To appear in: Geomorphology

Received date: 2 June 2013

Revised date: 12 September 2013
Accepted date: 16 September 2013

Please cite this article as: Moosavi, Vahid, Talebi, Ali, Shirmohammadi, Bagher, Produc-
ing a landslide inventory map using pixel-based and object-oriented approaches optimized
by Taguchi method, Geomorphology (2013), doi: 10.1016/j.geomorph.2013.09.012

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Producing a landslide inventory map using pixel-based and object-oriented

approaches optimized by Taguchi method
Vahid Moosavia, Ali Talebi*,a, Bagher Shirmohammadib

PT
a
Department of Watershed Management Engineering, Faculty of Natural Resources, Yazd University, Iran
b
Department of Environmental Science, Faculty of Natural Resources and Marine Sciences, Tarbiat

RI
Modares University, Iran

SC
NU
MA
D
P TE
CE
AC

* Corresponding author- Tel. & Fax: +98 351 8210312; E-mail: [email protected].
 ACCEPTED MANUSCRIPT

Abstract

Landslides are considered one of the most important natural hazards. Mapping landslides

and producing landslide inventory maps have received special attention from a wide

PT
range of specialists. The main objective of this study was to produce landslide inventory

RI
maps using advanced pixel-based (ANN and SVM) and object-oriented approaches. The

most important challenge in this case is to determine the optimum structure of

SC
classification methods. The Taguchi method was to perform optimization of the structure

NU
of ANN and SVM and segmentation process in the object-oriented classification method.

Results showed that the Taguchi method can be effectively used to cope with this
MA
problem. It significantly reduces the number of classification tests. We also showed that

there were no significant differences existed between ANN and SVM approaches ( 
2
D
TE

value of 3.33). However, we demonstrated that object-oriented approaches significantly

outperformed the pixel-based classification methods (Z value of 5.70) in producing a

P
CE

landslide inventory map. The accurate map produced using an object-oriented approach

(overall accuracy of 0.90) effectively determines the shape of landslides and also
AC

efficiently shows the intensifying effects of land use changes in the occurrence of

landslides.

Keywords: landslide inventory; Taguchi method; optimization; pixel-based classification;

object-oriented classification

1. Introduction

A landslide is the movement of a mass of rock, debris, or earth down a slope, under

the influence of gravity. Different phenomena can cause landslides, consisting of intense
 ACCEPTED MANUSCRIPT

or extended rainfall, swift snow melting, earthquakes (Wilson and Keefer, 1985; Inglès et

al., 2006), volcanic activity (Ablay and Hürliman, 2000; Moon and Simpson, 2002), and

human activities (Skempton and Hutchinson, 1969). Landslides can involve flowing,

PT
sliding, toppling, or falling; and many landslides exhibit a combination of two or more

RI
types of movements at the same time or during the lifetime of a landslide (Cruden and

Varnes, 1996; Philip and Ritz, 1999; Brückl, 2001; Sörensen and Bauer, 2003).

SC
Landslides are the most important, costly, and damaging natural hazards in mountainous

NU
regions (University of Utah, 1984). Landslide susceptibility analysis and mapping is a

key step in better predicting future landslides and improving protection in these areas.
MA
Several studies have been performed on landslides such as Schicker and Moon (2012)

that compared bivariate and multivariate statistical approaches for landslide susceptibility
D
TE

mapping or Pourghasemi et al. (2012) that produced a landslide susceptibility map using

the index of entropy. The use of GIS and remote sensing approaches to study landslides
P

are also intensively reported (Weirich and Blesius, 2007; Akgün et al., 2008; Booth et al.,
CE

2009; Martha et al., 2010; Pradhan and Lee, 2010; Oh et al, 2012; Song et al., 2012).
AC

The situations in which landslides are likely to occur are assumed to be the same

areas in which historical landslides have already occurred. Therefore, landslide inventory

mapping is a vital requirement for trustworthy hazard and risk analysis. Inventory maps

are available in only a few countries and mostly for limited areas (Ayenew and Barbieri,

2005; Duman et al., 2005; Harp et al., 2011). These maps can be used in several scientific

studies. Blahut et al. (2010) analyzed the landslide inventory maps to accurately predict

debris flow source areas. A combination of aerial photograph visual interpretation and

field work has conventionally been used to produce landslide inventories and, until now,
 ACCEPTED MANUSCRIPT

it remains the most often used approach for the extraction of inventory maps in scientific

studies and by managerial organizations (Hervas and Bobrowsky, 2009). Although this

approach is time consuming and labor intensive, the results are somewhat unreliable

PT
(Galli et al., 2008). Satellite image processing techniques can be suitable to produce more

RI
reliable inventory maps. There are several image processing techniques that can be

categorized into two groups: i.e., pixel-based and object-oriented approaches. The

SC
traditional digital image analysis approaches, which exclusively gain statistical methods,

NU
have proved to be constrained for detecting targets of greater complexity.

Pixel-based techniques classify each pixel in the image without regard to neighboring
MA
pixels. Several studies have been carried out using a pixel-based approach, and they have

examined several pixel-based techniques (Breiman, 2001; Huang et al., 2002; Pal, 2005;
D
TE

Carreiras et al., 2006; Gislason et al., 2006; Brenning, 2009; Otukei and Blaschke, 2010).

A number of pixel-based approaches are on hand for image classification, such as

P

maximum likelihood, minimum distance, parallelepiped, ISODATA, K-mean, etc.

CE

(Lillesand et al., 2007). Classification using pixel-based approaches has some

AC

deficiencies, especially dealing with the rich information content of high resolution data,

e.g., Geoeye multispectral (VNIR) and very high resolution (VHR) satellite imageries. In

fact, these conventional pixel-based approaches use only gray values; but the advanced

pixel-based techniques such as artificial neural network (ANN) and support vector

machine (SVM) consider the texture, tone, and several other characteristics (Pakhale and

Gupta, 2010). A neural network is, like supervised classification, a method that is first

trained from known data and then used to categorize unknown pixels. Support vector

machines (SVMs) demonstrate a set of theoretically superior machine-learning

 ACCEPTED MANUSCRIPT

algorithms. The development of SVM was initially triggered by the exploration and

formalization of learning machine capacity control and overfitting issues (Vapnik, 1998).

Object-oriented approaches are also one of the most useful image processing

PT
techniques. These approaches are based on the assumption that a pixel is very likely to

RI
belong to the same class as its neighboring pixels. In the first step, the image is

segmented into homogeneous objects consisting of similar pixels. These objects are then

SC
categorized into specific classes (Koch et al., 2003). Object-based image analysis have

NU
been applied in various fields in environmental sciences and disaster monitoring and

assessing (Langanke et al., 2004; Blaschke, 2005, 2010; Chandra et al., 2005; Laliberte et
MA
al., 2006; Castillejo-González et al., 2009; Linke and McDermid, 2011; Van Den

Eeckhaut et al., 2012).

D
TE

One of the most important challenges in these cases is the determination of the

optimum combination of affecting parameters on the performance of classification

P

approaches. Trial and error approaches are usually time consuming and labor intensive. A
CE

fractional factorial design of experiments such as the Taguchi method can be an effective
AC

way to cope with this problem. Taguchi (1990) developed a family of FFE matrices that

could be utilized in various situations. This method has been generally adopted to

optimize the design parameters (based on a signal-to-noise parameter) and significantly

minimize the overall testing time and the experimental costs (Wang and Huang, 2007;

Chou et al., 2009) following a systematic approach to confine the number of experiments

and tests.

To the best of our knowledge, no research uses the Taguchi method to optimize the

classification parameters. Several researchers compared the two classification approaches

 ACCEPTED MANUSCRIPT

(Yan et al., 2006; Platt and Rapoza, 2008; Castillejo-González et al., 2009; Myint et al.,

2011); but a few researchers evaluated the statistical significance of differences between

these two classification approaches, especially in producing landslide inventory maps

PT
(Pakhale and Gupta, 2010). The main objective of this research is to produce a landslide

RI
inventory map using Taguchi-based optimized advanced pixel-based and object-oriented

approaches and also to compare these methods using statistical indices.

SC
NU
2. Study area

The study site is located 40 km southeast of Kermanshah city, Kermanshah Province,

MA
Iran, between 47°00´ to 47°07´ E. longitude and 33°40´ to 34°00´ N. latitude. The

minimum and maximum elevations in this region are 1580 and 1950 m asl, respectively.
D
TE

This area encompasses plains, hilly, and mountainous areas with forest, rangeland, and

agricultural areas. The mean annual precipitation and temperature are 481 mm and
P

17.7°C, respectively. This area is vulnerable to landslide, which is partly related to its
CE

geological origin. This area is almost completely occupied by the Zagros fold and thrust
AC

belt (Zagros FTB); the Zagros FTB is due to the oblique convergence with the Arabian

plate and the Eurasian plate at 3 cm/y. Deformation within the Zagros FTB began during

the Oligocene and continues until the present day. This region is formed by a sequence of

Precambrian–Pliocene shelf sediments (Karimibavandpoor et al., 1999) and contains

many evaporitic layers (Alavi, 2004) that are a good candidate for landslide shear plane.

The other reasons for this high vulnerability can be related to the uncontrolled

deforestation, overgrazing, and offensive tillage activities (Heshmati et al., 2011). Figure
 ACCEPTED MANUSCRIPT

1 shows the location of the study area and the main geological structures of the Zagros

FTB.

Fig. 1

PT
The Geoeye multispectral VNIR imagery (4 bands with spatial resolution of 1.65 m)

RI
for 29 March 2011 was used in this study. The spacecraft is intended for a sun-

synchronous orbit at an altitude of 681 km and an inclination of 98°, with a 10:30 a.m.

SC
Equator crossing time. Geo images are shipped with the sensor camera model in RPC

NU
format and a metadata file (GeoEye, 2009).
MA
3. Materials and methods

3.1. Image classification

D

Classification of images involves using a procedure to decide whether different

TE

pixels in an image have analogous characteristics. These rules segregate the total data
P

space into subsets divided by so-called decision boundaries. All pixels that fall within a
CE

number of pixels delimited by such decision borders are then labeled as belonging to a
AC

distinct class (Elachi and van Zyl, 2006). As mentioned before, we use two main

classification approaches: namely, pixel-based and object-oriented. Before classification,

the image was geometrically corrected with first-degree polynomial using homogenously

distributed ground control points (45 points) obtained from field surveys. This algorithm

was applied and achieved a root mean square error (RMSE) of 1.8 m. A brief explanation

of these two classification approaches are presented below.

3.1.1. Pixel-based approach

 ACCEPTED MANUSCRIPT

In this part, three essential steps were conducted, i.e., selection of training samples

that are representative for different information classes; executing classification

algorithms; and finally, assessing the accuracy of the classified images through analysis

PT
of a confusion matrix (Tso and Mather, 2009). Training samples were selected according

RI
to the ground truth data. These homogenous areas were identified in the image to form

the training samples for all of the information classes. Two advanced supervised pixel-

SC
based classifications, i.e., ANN and SVM, were conducted in this part.

NU
The advantage of neural networks is related to the high computation rate

accomplished by their inherent parallelism that is a result of a potent arrangement of

MA
interconnections (weights) and simple processors (neurons), which allows processing of

very large data sets. This approach is commonly described as nonparametric (Frizzelle
D
TE

and Moody, 2001). The revenue of a neural network depends on how appropriate it has

been trained. During the training phase, the neural network learns about regularities
P

present in the training data and, based on these regularities, constructs rules that can be
CE

extended to the unknown data. This is one particular ability of neural networks (Tso and
AC

Mather, 2009).

In the neural network classification, the most accepted algorithm commonly used for

updating the neuronal activities and the interconnection in a multilayer perceptron (i.e.,

back-propagation algorithm) was used for supervised classification of images using the

ENVI software package. Back-propagation consists of two main steps, i.e., forward and

backward propagation, to achieve its adjustment of the neural state. In this approach,

learning takes place by regulating the weights in the node to minimize the difference

between the output node activation and the desired output. The error is back propagated
 ACCEPTED MANUSCRIPT

through the network, and weight modification is made using a recursive method

(Hopfield, 1982; Richards, 1999). In both approaches (ANN and SVM), in addition to the

original image, NDVI, a digital elevation model (DEM) and some of its derivatives (i.e.,

PT
slope and curvature) have been used. Figure 2 shows the original image and examples of

RI
deforestation and landslides.

Fig. 2

SC
The SVM also is a classification system resulting from statistical learning theory that

NU
provides good classification results from complex data. There are four main kernel types

in SVM: namely, linear, polynomial, radial basis function, and sigmoid. All of these are
MA
different ways of mathematically representing a kernel function (Hsu et al., 2007). This

approach is a binary classifier in which n-class problems can be transformed into the
D
TE

sequence of n binary classification tasks (Belousov et al., 2002). The SVM differs from

other separating hyperplane approaches in the way the hyperplane is constructed from the
P

training points (Marjanović et al., 2011). Figure 3 shows a linear SVM as an example that
CE

used a linear kernel defining the SVM hyperplanes. The data close to the hyperplane
AC

defined the support vectors of this hyperplane. This method used a penalty parameter that

allows identifying misclassification observed in the input data set. In this study, all four

abovementioned kernel types were tested.

Fig 3

Both of the mentioned classification approaches have several parameters that should

be tuned and optimized. Training threshold contribution, training rate, training

momentum, minimum output activation threshold, and the number of iterations are the

parameters that should be optimized in the ANN-based classification method. In the

 ACCEPTED MANUSCRIPT

SVM-based classification method kernel function, gamma, penalty parameter, and

pyramid level should be optimized. The Taguchi-based optimization process of the

mentioned parameters is explained in section 3.2.

PT
RI
3.1.2. Object-oriented approach

In this part, similar to pixelbased approaches, in addition to the original image,

SC
NDVI, DEM, and some of its derivatives (i.e., slope and curvature) have been used.

NU
Object-oriented approaches are based on two steps: segmentation and classification.

Segmentation or dividing images into homogeneous areas (objects) decreases the inherent
MA
noise in pixel-based analysis and facilitates a multiscale analysis (Blaschke, 2010). Rule-

based classification of those segments, incorporate spectral, textural, morphometric, and

D

contextual features. Obviously, the quality of the segmentation regularly controls the
TE

classification. Therefore, while pixel-based methods first classify pixels, considering

P

them as spatially independent bodies to subsequently define the object boundaries, the
CE

object-oriented method performs more reasonably by first delineating and then

AC

classifying objects (Draguţ and Eisank, 2011).

Image segmentation is a fundamental step for the object-oriented automatic feature

extraction procedure. It is a spatial clustering technique that entirely subdivides the image

into nonoverlapping areas or segments. There are two main segmentation categories:

namely edge-based and region-based methods. In the edge-based algorithms, image edges

are detected by thresholding of the image gradient obtained from a differentiation filter;

while in the region-based methods a group of pixels is detected either by growing a ‘seed’

pixel until the fulfillment of the homogeneity criteria or by splitting the entire image

initially and then merging the homogeneous regions (Möller et al., 2007). In this study a
 ACCEPTED MANUSCRIPT

multiresolution algorithm was used in the segmentation process. This algorithm is a

bottom-up region-merging technique starting with one pixel that continues in several

steps until the fulfillment of a user-defined homogeneity criterion (Benz et al., 2004).

PT
Multiresolution segmentation is controlled by several factors: namely, scale, color

RI
(color/shape ratio), and shape (compactness/smoothness ratio) parameters. The optimum

combination of these parameters was determined using Taguchi method (section 3.2).

SC
In this study, eCognition software was used for object-oriented classification. The

NU
approach for recognition and classification of information classes (such as landslides,

road, forest, etc.) is chiefly obtained from the knowledge developed by specialists for
MA
detection of these classes during image interpretation. Therefore, it is very similar to the

cognitive approach that is used by a specialist in visual image interpretation (Navulur,

D
TE

2007). When objects are accurately delineated (after optimizing the segmentation

parameters), eCognition computes an immense number of parameters, such as shape,

P
CE

texture, relationship, and layer mean for each object— to be used as a class discriminator

in the object-oriented classification step. These attributes of image objects are referred to
AC

as object features. So, in order to properly label the obtained objects, the best metrics and

characteristics of different information classes were assessed using two approaches: i.e.,

expert knowledge and statistical assessment of the metrics and characteristics for each

class. eCognition has a data mining functionality called ‘feature space optimization’ that

allows users to identify the best characteristics of objects (e.g., shape, texture,

relationship, layer mean, etc.) to be used for classification, based on training sets. Feature

space optimization was used to find the combination of features that are appropriate for
 ACCEPTED MANUSCRIPT

separating classes. In the last step, a nearest neighbor algorithm was performed to classify

the satellite imagery.

PT
3.2. Taguchi-based optimization of classification parameters

RI
Determination of the optimum value of the mentioned parameters in all three

methods using a trial and error approach is typically time consuming and labor intensive.

SC
Therefore, using an optimization method can be a good idea to cope with this problem.

NU
Several approaches are used by researchers to optimize object-oriented classification

parameters (Laliberte et al., 2004; Barlow et al., 2006; Dragut and Blaschke, 2006; Kerle
MA
and de Leeuw, 2009). However, most of these approaches just optimize the scale not the

combination of all mentioned parameters. In this study, a robust statistical approach,

D

namely the Taguchi method (Taguchi, 1990), was used to determine the optimum
TE

combination of classification and segmentation parameters. The effect of many different

P

parameters on the performance attribute in a reduced set of experiments can be examined

CE

using Taguchi’s orthogonal array experimental design. Taguchi’s orthogonal arrays offer
AC

an alternative to standard factorial designs. An orthogonal array is a type of experiment

where the columns for the independent variables are ‘orthogonal’ to one another. The use

of these tables makes the design of experiments very easy and consistent. Taguchi

orthogonal array designs are often used in design experiments with multiple level factors.

They can be thought of as a general fractional factorial design.

In this step, the kappa coefficients in the pixel-based approaches and Espindola et

al.’s (2006) objective function (namely plateau objective function or POF) in the

segmentation process (for object-oriented approach) were maximized. The kappa

coefficient is a well known index, and POF is a combination of a spatial autocorrelation

 ACCEPTED MANUSCRIPT

index (e.g., Moran’s I index) and a variance indicator to assess the segmentation quality

and to find optimal segmentation parameters.

In the SVM classification method, an initial test showed that linear and sigmoid

PT
kernel functions did not produce acceptable results. It may be because of the type of

RI
kernel function that could not classify and cluster the pixel values (DNs) accurately.

Thus, the confusion matrixes obtained from these classifications showed unpleasant

SC
results. Therefore, these kernel functions were eliminated from more analysis. In the

NU
object-oriented approach, an initial test also showed that scale parameters > 50, tended to

undersegment the image with noticeable mixes of objects. Scale parameters < 10 also
MA
tended to oversegment the image with many adjacent objects of the same land cover

observed. We considered five levels for each of scale, color, and shape parameters. The
D
TE

most important factors and levels can be detected using literature review as well as the

scientific basis of the method. For example, from a technical view it is obvious that the
P

activation function or the numbers of iterations are of the most important parameters in
CE

the ANN structure. Table 1 shows the parameters that affect the performance of ANN and
AC

SVM and also segmentation approaches. According to the number of factors and their

levels, the appropriate Taguchi orthogonal arrays were determined. An L32 orthogonal

array was used to optimize ANN and SVM parameters and a L25 orthogonal array were

used to optimize segmentation parameters. Numbers 1 to 5 show the levels of each

parameter. In fact, Taguchi minimizes the number of tests using standard orthogonal

arrays. For example, L25 (35) corresponds to an orthogonal array of three parameters,

each of which has five levels, and offers just 25 tests instead of 243 tests that is

mandatory in a full factorial design of experiment.

 ACCEPTED MANUSCRIPT

In the next step, classification and segmentation tests were performed according to

the selected orthogonal arrays. In the ANN and SVM classification tests, the cappa

coefficient, and in the object-oriented classification tests, the POF were calculated for

PT
each of the tests specified in the orthogonal arrays.

RI
Table 1.

Then, an analysis of the signal-to-noise (S/N) ratio was used to evaluate the

SC
classification and segmentation results. As this study aimed to maximize the kappa

NU
coefficient and POF, the S/N ratio with ‘higher is better’ (HB) characteristics was

selected for the study rather than the two other types of S/N ratio analyses, i.e., ‘lower is
MA
better’ (LB) and ‘nominal is best’ (NB). The S/N ratio for the HB type was then

calculated based on the following equation:

D

1 1 
TE

SNR  10Log10   2  (1)

n yi 
P

where n is the number of repetitions under the same experimental conditions (i.e., 1 in
CE

this study), and y represents the result of measurement. Here, y is the kappa coefficient
AC

for ANN, and SVM, and POF values obtained from each segmentation test. The S/N

response table and figure were then obtained, and the optimal conditions were

recognized. As a final point, the confirmation tests under these optimal conditions were

carried out.

4. Accuracy assessment

In order to assess accuracy, confusion matrixes were used for pixel-based and object-

oriented approaches. A confusion matrix is a square array of dimension r × r, where r is

the number of classes. The matrix shows the relationship between two samples of
 ACCEPTED MANUSCRIPT

measurements taken from the area that has been classified. The evaluation of the

statistical significance of the difference in accuracy between two classified images has

often been based on the comparison of the kappa coefficient calculated for each map. The

PT
kappa coefficient was then calculated using Eq. (2). The significance of the difference in

RI
accuracy between two maps with independent kappa coefficients was evaluated using

Eqs. (3) to (8) (Foody, 2004):

SC
 
Kˆ  1 2 (2)
1  2

NU
kˆ1  kˆ2
z
MA (3)
ˆ 2 k  ˆ 2 k
1 2

ˆ 
2 ˆ
 k  
1 1 1  1  21  1 21 2   3  1  1   4  4 2 
 
2 2


 
D

r  1   2 2 1   2 3 1   2 4  (4)
TE

The variables 1 ,  2 ,  3 , and  4 are calculated as below:

P

1 r r
1   ii 
CE

x  pii (5)
n i 1 i 1
AC

1 r r
2   i  i 
n 2 i 1
x  x 
i 1
pi   p i (6)

r r

 xii  xi   xi   pii   pi   pi 

1
3  (7)
n2 i 1 i 1

1 r r
  pij   p i  p j  
r r

3  ij 
4    
2 2
x xi x j (8)
n i 1 j 1 i 1 j 1

where, K̂ is the kappa coefficient, r is the number of columns (and rows) in a confusion

matrix, xii is entry (i, i) of the confusion matrix, xi+ and x+i are the marginal totals of row i

and column j, respectively, and n is the total number of observations. The overall
 ACCEPTED MANUSCRIPT

accuracy is obtained by dividing the sum of the main diagonal entries of the confusion

matrix by the total number of samples. The significance of the difference between the two

kappa coefficients is then assessed by comparing the value of Z calculated from Eq. (2)

PT
against tabulated values. As pixel-based and object-oriented approaches have different

RI
samples (i.e., independent kappa coefficients), Z statistical index was used to compare

results between these two approaches. The hypothesis that two kappa coefficients are

SC
equal is rejected if Z > 1.96 (95% confidence level).

NU
The McNemar test that is based upon the standardized normal test statistic can be

used to compare two related kappa coefficients (Foody, 2004; Zar, 2009):
MA
( f12  f 21 ) 2
2  (9)
f12  f 21
D

where fij indicates the frequency of sites lying in confusion matrix elements i, j (Bradley,
TE

1968). As two pixel-based approaches have the same samples (i.e., related kappa
P

coefficients), this statistical index was used to compare results between them. The
CE

hypothesis that two kappa coefficients are equal is rejected if  > 3.84 (95% confidence
2
AC

level). The mentioned certain values are obtained from Z and  statistical tables.
2

5. Results and discussion

5.1. Optimization results

As already explained, 32 classification prototypes were tested for each of ANN an

SVM approaches according to the Taguchi orthogonal array. The L32 (54) orthogonal

array and the value of the kappa coefficient obtained through classification tests for both

approaches have been presented in Table 2. Table 3 and Fig. 4 show the values of the S/N
 ACCEPTED MANUSCRIPT

ratio and mean effect plot for the ANN and SVM approaches. The boldface figure refers

to the maximum value of the S/N ratios of a certain factor among four levels, and thus it

indicates the optimum conditions for classification. As Table 3 and Fig. 4 show, the

PT
optimum conditions for the ANN approach are as follows: (i) activation function:

RI
hyperbolic; (ii) training rate: 0.1; (iii) training momentum: 0.7; (iv) Iterations: 5000; and

(v) minimum output activation threshold: 1.00e-9. The optimum conditions for the SVM

SC
approach are also as follows: (i) kernel function: RBF; (ii) gamma: 0.3; (iii) penalty

NU
parameter: 100; (iv) pyramid levels: 1; and (v) pyramid reclassification threshold: 0.9.

Table 2.
MA
Table 3.

Fig. 4
D
TE

Table 4 also shows the L25 and value of POF obtained through segmentation tests.

The values of the S/N ratio and the mean effect plot for POF are shown in Table 5 and
P

Fig. 5. The boldface figure refers to the maximum value of the S/N ratios of a certain
CE

factor among four levels, and thus it indicates the optimum conditions for segmentation.
AC

As Table 5 and Fig. 5 show, the optimum conditions for segmentation are as follows: (i)

the scale of 50; (ii) the color parameter equal to 0.3; and (iii) the shape parameter equal to

0.5. An important point in the object-oriented approach is that a feature of variable sizes,

such as landslide, typically cannot be delineated with a single-scale parameter. Therefore,

instead of selecting a single-scale factor with the highest objective function value, two

other scales with the highest POF values within the tested segmentations were also

chosen. As mentioned before, the color parameter (color/shape ratio) is equal to 0.3,
 ACCEPTED MANUSCRIPT

which demonstrates that the shape has a great importance in the segmentation process in

the mentioned image.

Table 4.

PT
Table 5.

RI
Fig.5

5.2. Classification and accuracy assessment results

SC
As the pixel-based classified images often suffer from a lack of spatial coherency

NU
(speckle or holes in classified areas), clumping and generalization were performed to

smooth them and to eliminate the few isolated pixels that did not have geomorphological
MA
significance. The selected classes were clumped together by first performing a dilate

operation and then an erode operation on the classified image using a 3×3 kernel. In the
D

next step, all other objects have been merged because the objective of this study was to
TE

produce landslide inventory maps. Figures 6 and 7 show the best pixel-based
P

classification results and final inventory maps.

CE

Fig. 6
AC

Fig. 7

Table 6 shows the summary of confusion matrices for ANN, SVM, and object-

oriented approaches. According to Figs. 6 and 7 and Table 6, both pixel-based

approaches have an approximately poor performance and several misclassifications

happened in them. This poor performance may be related to the high spectral variance in

the high resolution imagery. As in the pixel-based approaches, only the spectral

characteristics of the phenomena are considered in the classification process; it cannot

cope with the high spectral variance in each phenomenon and spectral similarities

between different phenomena. The other important point that can be deduced from these
 ACCEPTED MANUSCRIPT

tables is the superior performance of SVM in comparison with the ANN approach for

landslide detection (for linear structures, like road and rivers, it is not the case).

The McNemar test on the other hand, shows that there is not a statistically significant

PT
difference between ANN and SVM methods with a  value of 3.33 (insignificant at the
2

RI
95% confidence level). This result is in contrast with Pakhale and Gupta (2010) who

SC
showed that ANN outperformed the SVM method in classifying ETM+ data. It may be

related to the differences in the nature of the data used in these two studies. One of the

NU
main differences between these two images can be the spatial resolution.
MA Table 6

Figure 8 shows the object-oriented classification results and final inventory map. As
D

it is obvious in this map, several landslides have occurred adjacent to the farmlands,
TE

especially where the forest is removed and land use has been changed to agriculture. It

shows the intensifying effect of agriculture on mass movement, particularly on hillslopes

P
CE

that are accompanied with deforestation. Shili and Marn layers and the Goorpi Formation

can also be a triggering factor for landslides. Shili and Marn layers and the Goorpi
AC

Formation include a great amount of clay. Clay, a weathering product of rock mass,

contributes to landslide occurrence because of their chemical and physical properties.

Low permeability and clay-rich soils can trigger/reactivate landslides. The phenomena of

slope instability and landslides are particularly caused by the shrink – swell cycles that

occur in clay soils. The pore pressure in the Marn from groundwater can also be one of

the important triggering factors for landslides. The amount of precipitation is the other

important factor; when it reaches the terrain surface, it can complete the segmentation

rules for landslide bodies.

 ACCEPTED MANUSCRIPT

The statistical comparison between object-oriented and pixel-based approaches

demonstrates a significant difference between these two methods with the Z statistics

value of 5.70 (significant at the 95% confidence level). The superior performance of the

PT
object-oriented classification method is due to considering the geometric properties of the

RI
features in addition to its spectral properties.

Fig. 8

SC
The improved classification using object-oriented algorithms may be related to the

NU
use of objects (instead of pixels) to decrease the spectral variance in each information

class, especially in high resolution imagery. Also, clumping and generalization can
MA
improve the appearance of pixel-based classification regarding isolated misclassified

pixels, but the spectral heterogeneity in each of the information classes and spectral
D
TE

similarities between different phenomena are still challenging. For example, within the

pixel-based classification, many of the pixels classified as landslides are actually sparse
P

forest. So the number and area of landslides are overestimated in this approach. Object-
CE

oriented approaches can resolve some of these problems encountered using pixel-based
AC

methods to produce landslide inventory maps that are very important and useful for

managerial objectives. One of the advantages of object-oriented classification is the

utility of contextual information related to objects such as geometric characteristics in

addition to spectral characteristics and auxiliary data, e.g., DEM and its derivatives. This

result is in accordance with the results of Jobin et al. (2008) who noted that one of the

advantages of the object-oriented approach is the utility of a knowledge base that is

beyond purely spectral information and incorporating object-related features such as

 ACCEPTED MANUSCRIPT

shape, texture, and context/relationship, along with the capability to include ancillary

data.

6. Conclusion

PT
The main goal of this research was to provide an inventory map of landslides more

RI
precise using two powerful approaches, i.e., object-oriented image classification and the

Taguchi method. While pixel-based classification techniques process each pixel

SC
independently without taking into account information about spatial structures, further

NU
improvement of classification performances can be achieved by the incorporation of

spatial information into a classifier, especially in areas where structural information is

MA
important to distinguish between classes. Although object-oriented classifications are

comparatively more time consuming to produce than their pixel-based counterparts,

D
TE

object-oriented approaches can deal with the mentioned problems and challenges. The

results show that the object oriented approaches are not very sensitive to noise of the
P

image and avoids landslide misclassifications. The results of this study also show a
CE

statistically significant difference in the accuracy between the pixel-based SVM and the
AC

ANN approaches for landslide characterization and the object-oriented nearest-neighbor

classifier for mapping landslides using Geoeye data. Consequentially, poor pixel-based

classification confirms that thematic mapping using high spatial resolution satellite

imagery necessitates a new methodology that gains spectral and spatial (geometric)

properties of different phenomena. This is where the object-oriented approach has a great

advantage over pixel-based classification methods. We also demonstrated that the

Taguchi method can be effectively used to optimize the structure of the classification
 ACCEPTED MANUSCRIPT

methods. Using Taguchi orthogonal arrays makes it very easy and consistent to find the

best combination of classification parameters performing a limited number of tests.

PT
References

RI
Ablay, G., Hürliman, M., 2000. Evolution of the north flank of Tenerife by

SC
recurrent giant landslides. Journal of Volcanology and Geothermal Research

103, 135–159.

NU
Akgün, A., Dag, S., Bulut, F., 2008. Landslide susceptibility mapping for a
MA
landslideprone area (Findikli, NE of Turkey) by likelihood-frequency ratio and

weighted linear combination models. Environmental Geology 54, 1127–1143.

D

Alavi, M., 2004. Regional stratigraphy of the Zagros fold and thrust belt of Iran
TE

and its proforeland evolution. American Journal of Science 3004, 1–20.

P

Ayenew, T., Barbieri, G., 2005. Inventory of landslides and susceptibility mapping
CE

in the Dessie area, northern Ethiopia. Engineering Geology 77, 1–15.

AC

Barlow, J., Franklin, S., and Martin, Y., 2006. High spatial resolution satellite

imagery, DEM derivatives, and image segmentation for the detection of mass

wasting processes. Photogrammetric Engineering and Remote Sensing 72(6),

687–692.

Belousov, A.I., Verzakov, S.A., Von Frese, J., 2002. Applicational aspects of

support vector machines. Journal of Chemometrics 16, 482–489.

Benz, U., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2004. Multi-

resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready

 ACCEPTED MANUSCRIPT

information. ISPRS Journal of Photogrammetry & Remote Sensing 58, 239–

258.

Blahut, J., van Weste, C.J., Sterlacchini, S., 2010. Analysis of landslide

PT
inventories for accurate prediction of debris-flow source areas.

RI
Geomorphology 119, 36–51.

SC
Blaschke, T., 2005. Towards a framework for change detection based on image

objects. In: Remote Sensing and GIS for Environmental Studies. Erasmi, S.,

NU
Cyffka, B., Kappas, M. (Eds.), Goettinger Geographische Abhandlungen,
MA
Goettingen, Germany, pp. 1–9.

Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS
D

Journal of Photogrammetry and Remote Sensing 65(1), 2–16.

TE

Booth, A.M., Roering, J.J., Perron, T., 2009. Automated landslide mapping using
P

spectral analysis and high-resolution topographic data: Puget Sound lowlands,

CE

Washington, and Portland Hills, Oregon. Geomorphology 109, 132–147.

AC

Bradley, J.V., 1968. Distribution-Free Statistical Tests. Prentice-Hall, Englewood

Cliffs, NJ, 388 pp.

Breiman, L., 2001. Random forests. Machine Learning 45, 5–32.

Brenning, A., 2009. Benchmarking classifiers to optimally integrate terrain

analysis and multispectral remote sensing in automatic rock glacier detection.

Remote Sensing of Environment 113(1), 239–247.

Brückl, E.P., 2001. Cause-effect models of large landslides. Natural Hazards 23,

291–314.
 ACCEPTED MANUSCRIPT

Carreiras, J.M.B., Pereira, J.M.C., Campagnolo, M.L., Shimabukuro, Y.E., 2006.

Assessing the extent of agriculture/pasture and secondary succession forest in

the Brazilian Legal Amazon using SPOT VEGETATION data. Remote

PT
Sensing of Environment 101(3), 283–298.

RI
Castillejo-González, I.L., López-Granados, F., García-Ferrer, A., Peña-Barragán,

SC
J.M., Jurado-Expósito, M., de la Orden, M.S., 2009. Object- and pixel-based

analysis for mapping crops and their agro-environmental associated measures

NU
using QuickBird imagery. Computers and Electronics in Agriculture 68(2),
MA
207–215.

Chandra, M., Moreira, A., Keydel, W., 2005. Amper: Network on applied
D

multiparameter environmental remote sensing. An EU sponsored research and

TE

training network. In: Proceedings of the IGARSS 2005 Symposium, Seoul,

P

Korea.
CE

Chou, C.S., Ho, C.Y., Huang, C.I., 2009. The optimum conditions for combination
AC

of magnetic particles driven by a rotating magnetic field using the Taguchi

method. Adv. Powder Technology 20, 55–61.

Cruden, D.M., Varnes, D.J., 1996. Landslide types and processes. In: Turner,

A.K., Schuster, R.L. (Eds.), Landslides, Investigation and Mitigation. Special

Report 247, Transportation Research Board, Washington, DC, pp. 36–75.

ISSN: 0360-859X, ISBN: 030906208X.

Dragut, L., Blaschke, T., 2006. Automated classification of landform elements

using object-based image analysis. Geomorphology 81, 330–344.

 ACCEPTED MANUSCRIPT

Dragut, L., Eisank, C., 2011. Object representations at multiple scales from digital

elevation models. Geomorphology 129, 183–189.

Duman, T.Y., Can, T., Emre, O., Kecer, M., Dogan, A., Ates, S., Durmaz, S.,

PT
2005. Landslide inventory of northwestern Anatolia, Turkey. Engineering

RI
Geology 77, 99–114.

SC
Elachi, Ch., van Zyl, J., 2006. Introduction to the Physics and Techniques of

Remote Sensing. John Wiley & Sons, Inc. Jahr 552 pp.

NU
Espindola, G.M., Camara, G., Reis, I.A., Bins, L.S., Monteiro, A.M., 2006.
MA
Parameter selection for region-growing image segmentation algorithms using

spatial autocorrelation. International Journal of Remote Sensing 27, 3035–

D

3040.
TE

Foody, G.M., 2004. Thematic map comparison: evaluating the statistical

P

significance of differences in classification accuracy. Photogrammetric

CE

Engineering and Remote Sensing 70, 627−633.

AC

Frizzelle, B.G., Moody, A., 2001. Mapping continuous distributions of land

cover— a comparison of maximum likelihood estimation and artificial neural

networks. Photogrammetric Engineering and Remote Sensing, Vol, 67693–

67705.

Galli, M., Ardizzone, F., Cardinali, M., Guzzetti, F., Reichenbach, P., 2008.

Comparing landslide inventory maps. Geomorphology 94, 268–289.

GeoEye, Inc., 2009. GeoEye Product Guide. http://www.geoeye.com/CorpSite/

assets/docs/brochures/GeoEye Product Guide.pdf (accessed 03.02.12).

 ACCEPTED MANUSCRIPT

Gislason, P.O., Benediktsson, J.A., Sveinsson, J.R., 2006. Random forests for land

cover classification. Pattern Recognition Letters 27(4), 294–300.

Harp, E.L., Keefer, D.K., Sat, H.P., Yagi, H., 2011. Landslide inventories: the

PT
essential part of seismic landslide hazard analyses. Engineering Geology 122,

RI
9–21.

SC
Hervas, J., Bobrowsky, P., 2009. Mapping: inventories, susceptibility, hazard and

risk. In: Sassa, K., Canuti, P., (Eds), Landslides — Disaster Risk Reduction.

NU
Springer, Berlin, pp. 321–349. MA
Heshmati, M., Arifin, A., Shamshuddin, J., Majid, N.M., Ghaituri, M., 2011.

Factors affecting landslides occurrence in agro-ecological zones in the Merek

D

catchment, Iran. Journal of Arid Environments 75, 1072–1082.

TE

Hopfield, J.J., 1982. Neural networks and physical systems with emergent
P

collective computational abilities. Proceedings of the National Academy of

CE

Sciences of the United States 79, 2554–2558.

AC

Hsu, C.W., Chang, C.C., Lin, C.J., 2007. A practical guide to support vector

classification. National Taiwan University, URL

http://ntu.csie.org/~cjlin/papers/guide/guide.pdf.

Huang, C., Davis, L.S., Townshend, J.R.G., 2002. An assessment of support

vector machines for land cover classification. International Journal of Remote

Sensing 23(4), 725–749.

 ACCEPTED MANUSCRIPT

Inglès, J., Darrozes, J., Soula, J.C., 2006. Effects of vertical component of ground

shaking on earthquake-induced landslide displacements using generalized

Newmark’s analysis. Engineering Geology 86, 134–147.

PT
Jobin, B., Labrecque, S., Grenier, M., Falardeau, G., 2008. Object-based

RI
classification as an alternative approach to the traditional pixel-based

SC
classification to identify potential habitat of the Grasshopper Sparrow.

Environmental Management 41, 20–31.

NU
Karimibavandpoor, A., Hajihosaini, A., Shahandi, M., 1999. Geological Map of
MA
Kermanshah, Iran (1:100,000, Series: 5458). Geological Survey Service Press,

Tehran, Iran.
D

Kerle, N., de Leeuw, J., 2009. Reviving legacy population maps with object-
TE

oriented image processing techniques. IEEE Transactions on Geoscience and

P

Remote Sensing 47, 2392–2402.

CE

Koch, B., Jochum, M., Ivitis, E., Dees, M., 2003. Pixelbasierte Klassifizierung im
AC

Vergleich and zur Erganzung zum objektbasierten Verfahren. Photogrammetrie

Fernerkundung Geoinformation 3, 195–204.

Laliberte, A.S., Rango, A., Havstad, K.M., Paris, J.F., Beck, R.F., McNeely, R.,

Gonzalez, A.L., 2004. Object-oriented image analysis for mapping shrub

encroachment from 1937 to 2003 in southern New Mexico. Remote Sensing of

Environment 93 (1–2), 198–210.

Laliberte, A.S., Koppa, J.S., Fredrickson, E.L., Rango, A., 2006. Comparison of

nearest neighbor and rule-based decision tree classification in an object-

 ACCEPTED MANUSCRIPT

oriented environment. In: IEEE International Geoscience and Remote Sensing

Symposium Proceedings, July 31-August 4, Denver, Co.

Langanke, T., Blaschke, T., Lang, S., 2004. An object-based GIS/remote sensing

PT
approach supporting monitoring tasks in European-wide nature conservation.

RI
In: Proceedings of the Mediterranean Conference on Earth Observation, First

SC
Mediterranean Conference on Earth Observation (Remote Sensing), 21–23

April, Belgrade, Serbia, pp. 245–252.

NU
Lillesand, T., Kiefer, R.W., Chipman, J., 2007. Remote Sensing and Image
MA
Interpretation. London Wiley, 197 pp.

Linke, J., McDermid, G.J., 2011. A conceptual model for multi-temporal

D

landscape monitoring in an object-based environment. Selected Topics in IEEE

TE

Journal of Applied Earth Observations and Remote Sensing 4(2), 265–271.

P

Marjanović, M., Kovačević, M., Bajat, B., Voženílek, V., 2011. Landslide
CE

susceptibility assessment using SVM machine learning algorithm. Engineering

AC

Geology 123, 225–234.

Martha, T.R., Kerle, N., Jetten, V., Van Westen, C.J., Kumar, K.V., 2010.

Characterising spectral, spatial and morphometric properties of landslides for

semi-automatic detection using object-oriented methods. Geomorphology 116

(1–2), 24–36.

Möller, M., Lymburner, L., Volk, M., 2007. The comparison index: a tool for

assessing the accuracy of image segmentation. International Journal of Applied

Earth Observation and Geoinformation 9 (3), 311–321.

 ACCEPTED MANUSCRIPT

Moon, V., Simpson, C.J., 2002. Large-scale mass wasting in ancient volcanic

material. Engineering Geology 64, 41–64.

Myint, S.W., Gober, P., Brazel, A., Grossman-Clarke, S., Weng, Q., 2011. Per-

PT
pixel vs. object-based classification of urban land cover extraction using high

RI
spatial resolution imagery. Remote Sensing of Environment 115(5), 1145–

SC
1161.

Navulur, K., 2007. Multispectral image analysis using the object-oriented

NU
Paradigm. CRC Press, Taylor & Francis Group, Taiwan 206 pp.
MA
Oh, H.J., Park, N.W., Lee, S.S., Lee, S., 2012. Extraction of landslide-related

factors from ASTER imagery and its application to landslide susceptibility

D

mapping. International Journal of Remote Sensing 33 (10), 3211–3231.

TE

Otukei, J.R., Blaschke, T., 2010. Land cover change assessment using decision
P

trees, support vector machines and maximum likelihood classification

CE

algorithms. International Journal of Applied Earth Observation and

AC

Geoinformation 12, 27–31.

Pakhale, G.K., Gupta, P.K., 2010. Comparison of advanced pixel based (ANN and

SVM) and object-oriented classification approaches using landsat-7 ETM+

data. International Journal of Engineering and Technology 2(4), 245–251.

Pal, M., 2005. Random forest classifier for remote sensing classification.

International Journal of Remote Sensing 26(1), 217–222.

Philip, H., Ritz, J.F., 1999. Gigantic paleolandslide associated with active faulting

along the Bogd fault (Gobi-Altay, Mongolia). Geology 27, 211–214.

 ACCEPTED MANUSCRIPT

Platt, R.V., Rapoza, L., 2008. An evaluation of an object-oriented paradigm for

land use/land cover classification. The Professional Geographer 60(1), 87–100.

Pourghasemi, H.R., Mohammady, M., Pradhan, B., 2012. Landslide susceptibility

PT
mapping using index of entropy and conditional probability models in GIS:

RI
Safarood Basin, Iran. Catena 97, 71–84.

SC
Pradhan, B., Lee, S., 2010. Delineation of landslide hazard areas on Penang Island,

Malaysia, by using frequency ratio, logistic regression, and artificial neural

NU
network models. Environmental Earth Sciences 60, 1037–1054.
MA
Richards, J.A., 1999. Remote Sensing Digital Image Analysis. Springer-Verlag,

Berlin, 240 pp.

D

Schicker, R., Moon, V., 2012. Comparison of bivariate and multivariate statistical
TE

approaches in landslide susceptibility mapping at a regional scale.

P

Geomorphology 161–162, 40–57.

CE

Skempton, A.V., Hutchinson, J.N., 1969. Stability of natural slopes and

AC

embankment foundations. In: State of Art Report, 7th Conf. Soil Mech. Found.

Eng., Mexico, pp. 291–335.

Song, K.Y., Oh, H.J., Choi, J., Park, I., Lee, C., Lee, S., 2012. Prediction of

landslide using ASTER imagery and data mining models. Advances in Space

Research 49, 978–993.

Sörensen, S-A., Bauer, B., 2003. On the dynamics of the Köfels sturzstrom.

Geomorphology 54, 11–19.

 ACCEPTED MANUSCRIPT

Taguchi, G., 1990. Introduction to Quality Engineering. McGraw-Hill, New York,

USA, 191 pp.

Tso, B., Mather, P., 2009. Classification Methods for Remotely Sensed Data. CRC

PT
Press, Taylor & Francis Group, Boca Raton, 367 pp.

RI
University of Utah, 1984. Flooding and Landslide in Utah – An Economic Impact

SC
Analysis. bur. Of Econ. Devel. And Utah office of Plan. Budget, Salt Lake

city. Bureau of Economic Development and Utah Office of Planning and

NU
Budget, Salt Lake City, UT, 123 pp.
MA
Van Den Eeckhaut, M., Kerle, N., Poesen, J., Hervás, J., 2012. Object-oriented

identification of forested landslides with derivatives of single pulse LiDAR

D

data. Geomorphology 173–174, 30–42.

TE

Vapnik, V., 1998. Statistical Learning Theory. Wiley-Interscience. Available at:

P

http://www.amazon.ca/exec/obidos/redirect?tag=citeulike0920&path=ASIN/04
CE

71030031 [July 28, 2011].

AC

Wang, T.Y., Huang, C.Y., 2007. Improving forecasting performance by

employing the Taguchi method. European Journal of Operational Research

176, 1052–1065.

Weirich, F., Blesius, L., 2007. Comparison of satellite and air photo based

landslide susceptibility maps. Geomorphology 87 (4), 352–364.

Wilson, R.C., Keefer, D.K., 1985. Predicting areal limits of earthquake-induced

landsliding. In: Ziony, J.I., (Ed.), Evaluating Earthquake Hazards in the Los
 ACCEPTED MANUSCRIPT

Angeles Region – An Earth-Science Perspective. U.S. Geological Survey

Professional Paper 1360, United States 316–345.

Yan, G., Mas, J.F., Maathuis, B.H.P., Xiangmin, Z., Van Dijk, P.M., 2006.

PT
Comparison of pixel-based and object-oriented image classification

RI
approaches— a case study in a coal fire area, Wuda, Inner Mongolia, China.

SC
International Journal of Remote Sensing 27, 4039–4055.

Zar, J. H., 2009. Biostatistical analysis. 5th ed. Prentice Hall. United States 576

NU
pp. MA
D
P TE
CE
AC
 ACCEPTED MANUSCRIPT

Figure captions

Fig. 1. The study area. (A) Kermanshah province. (B) Geologic map of Iran and the main

geological structures of Zagros FTB.

PT
Fig. 2. (A) The original image; (B) examples of deforestation parts; (C) examples of

RI
landslides.

Fig. 3. Linear support vector machine example.

SC
Fig. 4. Main effects plot (data means) for S/N ratios for study variables of (A) ANN and

NU
(B) SVM approaches.

Fig. 5. Main effects plot (data means) for S/N ratios for study variables of segmentation
MA
process.

Fig. 6. Results of ANN classification.

D
TE

Fig. 7. Results of SVM classification.

Fig. 8. Results of object-oriented classification.

P
CE
AC
 ACCEPTED MANUSCRIPT

Fig. 1.

a)

PT
RI
Kermanshah

SC
NU
MA
D
P TE
CE
AC
 ACCEPTED MANUSCRIPT

Fig. 2.

PT
A) B)

RI
SC
NU
C)
MA
D
P TE
CE
AC
 ACCEPTED MANUSCRIPT

PT
Fig. 3.

RI
SC
NU
MA
D
TE
P
CE
AC
 ACCEPTED MANUSCRIPT

PT
Fig. 4.

Main Effects Plot (data means) for SN ratios

RI
A B C

-4.50

SC
-4.75

-5.00
Mean of SN ratios

-5.25

NU
-5.50
1 2 1 2 3 4 1 2 3 4
D E

-4.50 MA
-4.75

-5.00

-5.25

-5.50 ANN
1 2 3 4 1 2 3 4
D
TE

Main Effects Plot (data means) for SN ratios

A B C
P

-3.5
CE

-4.0

-4.5
Mean of SN ratios

-5.0
AC

1 2 1 2 3 4 1 2 3 4
D E
-3.5

-4.0

-4.5

-5.0

1 2 3 4 1 2 3 4
SVM
 ACCEPTED MANUSCRIPT

Fig. 5.

PT
Main Effects Plot (data means) for SN ratios
A B
1

RI
0

-1
Mean of SN ratios

SC
-2

-3
1 2 3 4 5 1 2 3 4 5
C
1

NU
0

-1

-2
MA OOA
-3
1 2 3 4 5
D
P TE
CE
AC
 ACCEPTED MANUSCRIPT

Fig. 6.

PT
RI
SC
NU
MA
D
TE
P
CE
AC
 ACCEPTED MANUSCRIPT

Fig. 7.

PT
RI
SC
NU
MA
D
TE
P
CE
AC
 ACCEPTED MANUSCRIPT

Fig. 8.

PT
RI
SC
NU
MA
D
TE
P
CE
AC
 ACCEPTED MANUSCRIPT

Table 1
Factors and their levels used for optimization in pixel-based and object-oriented
approaches

PT
Factor Description Level 1 Level 2 Level 3 Level 4 Level 5
SVM approach
A Kernel function RBF Polynomial - - -

RI
B Gamma 0.1 0.3 0.5 0.7 -
C Penalty parameter 10 100 1000 10000 -

SC
D Pyramid levels 1 2 3 4 -
Pyramid reclassification
E 0.3 0.5 0.7 0.9 -
threshold

NU
ANN approach
A Activation function Logistic Hyperbolic - - -
B Training rate 0.1 0.3 0.5 0.7 -
C Training momentum 0.3 0.5 0.7 0.9 -
MA
D Iterations 500 1000 5000 10000 -
Minimum output activation
E 1.00e-3 1.00e-5 1.00e-7 1.00e-9 -
threshold
Segmentation
D

A Scale 20 30 40 50 60
B Color 0.1 0.3 0.5 0.7 0.9
TE

C Shape 0.1 0.3 0.5 0.7 0.9

P
CE
AC
 ACCEPTED MANUSCRIPT

Table 2

PT
L32 orthogonal array and kappa coefficients for ANN and SVM classification
approaches

RI
L32 (combination of different levels) Kappa(SVM) Kappa(ANN)

SC
Test 1 1 1 1 1 1 0.658 0.501
Test 2 1 1 2 2 2 0.645 0.580
Test 3 1 1 3 3 3 0.525 0.677
Test 4 1 1 4 4 4 0.501 0.613

NU
Test 5 1 2 1 1 2 0.646 0.505
Test 6 1 2 2 2 1 0.704 0.503
Test 7 1 2 3 3 4 0.679 0.567
Test 8 1 2 4 4 3 0.650 0.554
Test 9 1 3 1
MA
2 3 0.572 0.447
Test 10 1 3 2 1 4 0.671 0.593
Test 11 1 3 3 4 1 0.650 0.562
Test 12 1 3 4 3 2 0.546 0.548
Test 13 1 4 1 2 4 0.645 0.563
D

Test 14 1 4 2 1 3 0.632 0.571

Test 15 1 4 3 4 2 0.501 0.530
TE

Test 16 1 4 4 3 1 0.534 0.523

Test 17 2 1 1 4 1 0.642 0.639
Test 18 2 1 2 3 1 0.629 0.646
Test 19 2 1 3 2 3 0.519 0.556
P

Test 20 2 1 4 1 4 0.555 0.512

Test 21 2 2 1 4 2 0.600 0.604
CE

Test 22 2 2 2 3 1 0.638 0.612

Test 23 2 2 3 2 4 0.673 0.620
Test 24 2 2 4 1 3 0.694 0.638
Test 25 2 3 1 3 3 0.556 0.623
AC

Test 26 2 3 2 4 4 0.655 0.602

Test 27 2 3 3 1 1 0.654 0.641
Test 28 2 3 4 2 2 0.530 0.537
Test 29 2 4 1 3 4 0.658 0.633
Test 30 2 4 2 4 3 0.586 0.610
Test 31 2 4 3 1 2 0.505 0.589
Test 32 2 4 4 2 1 0.518 0.492
 ACCEPTED MANUSCRIPT

PT
Table 3

RI
The S/N ratio for factors in each level for SVM and ANN approaches

SC
Factors
S/N ratios

NU
Level A B C D E
SVM
1 -4.345 -4.789 -4.211 -4.171 -4.200
MA
2 -4.617 -3.681 -3.888 -4.562 -4.925
3 -4.483 -4.756 -4.610 -4.670
4 -4.972 -5.069 -4.580 -4.130
D

ANN
1 -5.192 -4.618 -5.034 -4.939 -5.098
TE

2 -4.504 -4.833 -4.607 -5.433 -4.946

3 -4.940 -4.565 -4.412 -4.720
P

4 -5.000 -5.186 -4.608 -4.628

CE
AC
 ACCEPTED MANUSCRIPT

PT
RI
SC
Table 4
L25 orthogonal array and POF for segmentation process (object-oriented
classification approaches)

NU
L25 (combination
Test No. Scale Color Shape POF
of different levels)
MA
Test 1 1 1 1 20 0.1 0.1 0.60
Test 2 1 2 2 20 0.3 0.3 0.90
Test 3 1 3 3 20 0.5 0.5 0.82
Test 4 1 4 4 20 0.7 0.7 0.71
Test 5 1 5 5 20 0.9 0.9 0.65
D

Test 6 2 1 2 30 0.1 0.3 0.81

Test 7 2 2 3 30 0.3 0.5 0.98
Test 8 2 3 4 30 0.5 0.7 0.89
TE

Test 9 2 4 5 30 0.7 0.9 0.91

Test 10 2 5 1 30 0.9 0.1 0.82
Test 11 3 1 3 40 0.1 0.5 0.98
Test 12 3 2 4 40 0.3 0.7 1.16
P

Test 13 3 3 5 40 0.5 0.9 0.93

Test 14 3 4 1 40 0.7 0.1 1.14
CE

Test 15 3 5 2 40 0.9 0.3 0.99

Test 16 4 1 4 50 0.1 0.7 1.10
Test 17 4 2 5 50 0.3 0.9 1.17
Test 18 4 3 1 50 0.5 0.1 0.98
AC

Test 19 4 4 2 50 0.7 0.3 1.13

Test 20 4 5 3 50 0.9 0.5 1.00
Test 21 5 1 5 60 0.1 0.9 0.92
Test 22 5 2 1 60 0.3 0.1 0.98
Test 23 5 3 2 60 0.5 0.3 0.76
Test 24 5 4 3 60 0.7 0.5 0.83
Test 25 5 5 4 60 0.9 0.7 0.63
 ACCEPTED MANUSCRIPT

PT
RI
SC
NU
Table 5
The S/N ratio for factors in each level for object-oriented approaches
MA
Factors
Object-oriented
S/N ratios
Level A B C
1 -2.7585 -1.2678 -1.0747
D

2 -1.1122 0.2774 -0.831

TE

3 0.3068 -1.1851 -0.7386

4 0.6155 -0.6426 -1.1766
5 -1.783 -1.9132 -0.9104
P
CE
AC
 ACCEPTED MANUSCRIPT

PT
Table 6
Summary of confusion matrices of the three classification approaches

RI
SC
ANN SVM Object-oriented
PA(%) UA(%) CK PA(%) UA(%) CK PA(%) UA(%) CK
River 48.84 70.00 0.43 58.14 75.76 0.53 90.48 93.83 0.90

NU
Forest 84.21 85.33 0.79 81.58 83.78 0.76 93.23 86.06 0.91
Landslide 77.36 70.69 0.72 81.13 78.18 0.77 93.14 91.79 0.91
Road 62.79 65.85 0.57 72.09 68.89 0.67 85.35 88.02 0.82
Farmland 67.69 61.97 0.58 73.44 67.14 0.66 90.95 95.26 0.89
MA
Residential 86.21 73.53 0.85 80.00 75.00 0.78 84.95 84.95 0.83
Kappa 0.65 0.69 0.88
OA 0.71 0.75 0.90
V(K) 0.000965 0.000895 0.000135
D
P TE
CE
AC
 ACCEPTED MANUSCRIPT

Highlights
 Taguchi method was used to determine the optimum classification structure.
 Object-oriented and pixel-based classification methods were compared using
statistical approaches.
 Landslide inventory maps were produced.

PT
RI
SC
NU
MA
D
P TE
CE
AC