See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/270426525

Predicting rainfall-induced landslide potential along a mountain road in

Taiwan

Article  in  Géotechnique · February 2011

DOI: 10.1680/geot.8.P.119.3740

CITATIONS READS

20 206

3 authors, including:

Hung-Jiun Liao Jianye Ching

National Taiwan University of Science and Technology National Taiwan University
66 PUBLICATIONS   551 CITATIONS    216 PUBLICATIONS   4,975 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Subset simulation View project

Bayesian methods View project

All content following this page was uploaded by Jianye Ching on 26 December 2015.

The user has requested enhancement of the downloaded file.

 Predicting Rainfall-induced Landslide Potential along a Mountain Road in Taiwan

Jianye Ching1, Hung-Jiun Liao2, and Jiun-Yu Lee3

Revision Prepared on June 2, 2009

Number of words: 7000 Number of Tables: 6 Number of Figures: 8

Key words: slopes, landslides, failure, statistical analysis

1
(Corresponding Author) MSc, PhD, Associate Professor at the National Taiwan University, Taipei, Taiwan. Tel:
886-2-33664328. Email: jyching@gmail.com.
2
MSc, DIC, PhD, Professor at the National Taiwan University of Science and Technology, Taipei, Taiwan.
3
MSc, National Taiwan University of Science and Technology, Taipei, Taiwan.

1
ABSTRACT

This paper gives a brief introduction on the background and current status of rainfall-induced

landslides along a mountain road (Route T-18) in Taiwan and presents a detailed analysis based

on the Gaussian Process model for predicting locations and occurrence times of future land-

slides using post Chi-Chi earthquake historical data. Based on inherent and man-made features

of failed and not-failed slopes, locations of possible future landslides due to rainfall along Route

T-18 are predicted. Together with historical rainfall data, a rainfall fragility graph is established.

The analysis results show that the Gaussian Process model is effective in predicting landslide

potentials and probabilities. Comparisons of the Gaussian Process analysis and the discriminant

function analysis are made, which show that the former outperforms the latter in many aspects.

The results are valuable for predicting where and when landslides would occur along Route T-18

in future heavy rainfalls.

INTRODUCTION

The island of Taiwan was formed by the collision action of Eurasia plate and Philippine sea plate.

It is relatively young in geological age. For the total area of 36000 km2, mountains cover more

than two thirds of Taiwan. The area percentages of mountains are about 28.40% for elevation of

0~100m, 39.48% for 100~1000m, 20.22% for 1000~2000m, 10.73% for 2000~3000m, and

1.17% for 3000~4000m, respectively. To accommodate a population of more than 23 million, a

large portion of the lower mountain area has already been intensively developed for cultivation

and tourism. To keep up with the development, an extensive mountain road network has been

built over the past decades.

The total length of roads with elevation above 100m in Taiwan is more than 67,000 km.

2
Some of them were built with high engineering standards, but a large number of them were built

with low engineering standards. Even worse, some were built along river valleys by the cut-slope

methods and have suffered from both river scouring problems on the down-slope and stability

problems on the up-slope. Therefore, landslides associated with different failure types are not

unusual along mountain roads during long periods of rainfall or torrential rain associated with

typhoons.

When a landslide occurs, it may cause traffic interruption, damage to vehicles and injure

personnel. Therefore, it is desirable to predict where and when landslides may happen to ensure

transportation safety for the public. Research on landslide prediction is not new. In forest science,

prediction of landslides after a chaparral fire was investigated by Kojan et al. (1972), log-

ging-related landslides were analyzed using discriminant function analysis by Rice et al. (1985),

Pillsbury (1976), and Furbish (1981). Landslides in grass and brush environments were studied

by Waltz (1971) and Rice and Foggin (1971) using discriminant function analysis. The use of

discriminant function analysis and logistic regression with a geographic information system (GIS)

for predicting landslides was also studied by Carrera et al. (1991,1995), Mark and Ellen (1995),

Gorsevski et al. (2000), and Santacana et al. (2003). Chung and Fabbri (1999) proposed the use

of the Bayesian approach for landslide potential prediction. Wen et al. (2001) and Liao et al.

(2001) employed the discriminant function analysis to predict landslide probabilities of slopes in

hillside communities.

Landslides along mountain roads are somewhat different from other types of landslides:

they are affected by numerous features, especially manmade features. In this study, Route T-18 in

central Taiwan was chosen to demonstrate the suitability of landslide prediction using a Gaussian

Process model. Two main questions are of concern: (a) Given the historical rainfall-induced

3
landslide data along the mountain road in Taiwan, where are the locations along the route with

high landslide potential in a future rainfall? (b) Given the historical landslide and rainfall data,

what are the landslide probabilities of the slopes along the route in a future heavy rainfall? The

former mainly concerns with the locations of future rainfall-induced landslides, while the latter is

concerned with the time of landslide occurrence in future rainfalls.

Note that the concept of probability is highlighted in this study due to the complexity of

slope stability problems: it can never be certain where and when a slope will fail because there

are numerous uncertainties involved. In this study, a full probabilistic analysis based on the

Gaussian Process model (MacKay 1998; William and Rasmussen 1996; Neal 1997; Gibbs 1997)

is presented to predict the landslide potentials and probabilities. This new model is versatile and

flexible. Compared to discriminant function analysis (Fisher 1936), the Gaussian Process model

possesses the desirable property that there is no need to assume the functional form for the dis-

criminant function, which is a highly nontrivial task for complicated problems. Moreover, the

performance of the Gaussian Process analysis is comparable to a neural network approach while

the required computational cost is much less. Last but not least, in this paper the term “probabil-

ity” is not used in the frequentist sense but in a degree-of-belief sense. Therefore, the calculated

probability should not be interpreted as the actual relative frequency. In fact, later in the paper

the term “potential” may be used to replace the term “probability” because the former does not

involve the concept of relative frequency.

The structure of the paper is as follows: First the background information on mountain

roads in Taiwan is introduced, including information for rainfalls and typhoons. Second, the

landslide database adopted for representing mountain roads is discussed in detail, including the

definitions of the features used for the analysis. Then the analysis for predicting landslide loca-

4
tions and occurrence times is described, and the analysis results are presented and compared with

those made by the discriminant function analysis. Finally, a brief discussion will be made.

BACKGROUND

Typhoons

In total, 52 typhoons attacked Taiwan between 1994 and 2004 with an average of 4.7 typhoons

annually. Among them, Herb, Toraji, Nari and Mindulle were the most devastating ones and

caused a large number of landslides in Taiwan. Figure 1 shows the accumulative rainfalls rec-

orded at a rainfall station nearby Route T-18 (Alishan mountain road) during the week of each of

the four typhoons. It ranges from 1994mm for Typhoon Herb to 414mm for Typhoon Nari.

Landslides caused by these typhoons are classified into three major categories: rock fall and slips

on the up-slope (Type I), road base failure (Type II), and washed away road base (Type III).

Among them, rock fall and slip failure at the up-slope is the dominant failure type (Figure 2).

To deal with Type-I landslides, the route maintenance division tends to implement various

types of slope stabilization measures, such as a tied-down RC grid structure and gravity retaining

walls to protect the failed up-slopes. It is of interest to find that the number of Type-I damage

incidents decreases from 89 (in Herb) to 37 (in Mindulle), although the accumulative rainfalls of

Herb (1994 mm) and Mindulle (1764 mm) were not much different from each other. It indicates

that the slope stabilization measures have successfully increased the stabilities of failed slopes,

i.e.: slopes where landslides occurred. Once the failed slopes were properly protected, the reoc-

currences of landslides can be effectively controlled for some period of time. Moreover, the cost

spent on emergency repairs (short-term stabilization) and to stabilize the failed slopes and road

bases (long-term stabilization) decreases rapidly from Herb to Toraji and Nari although the rain-

5
fall of Mindulle is close to that of Herb.

However, the stabilization cannot be carried out on all the slopes along the routes due to fi-

nancial constraints, and new landslides may very likely occur in slopes that have not yet failed,

i.e.: slopes where landslides were not observed during the typhoons, and where stabilization

measures are absent. Therefore, it is desirable to develop techniques to predict the locations and

occurrence times of new landslides. This is achieved through analyzing historical data gathered

from the slopes along the route, which will be explained in detail in the next section.

Geology

Route T-18 is nearby the Alishan mountain in central Taiwan. The geologic condition along the

route mainly consists of sandstone and interlayered sandstone/shale strata with various bedding

structures. The geologic strata can be classified into five main categories in terms of their geo-

logic ages (from young to old): (a) Zuolan stratum, (b) Jinswei shale, (c) Dawo sandstone and

chalk, (d) Quandaushan sandstone, and (e) Nanzuan stratum. Among them, Nanzuan stratum is

the most widely distributed strata which consists of a thick sandstone layer, sandstone/shale in-

terlayer, and shale layer. About 90% of outcrops found along the route belong to the Nanzuan

stratum. The Zuolan stratum consists of thick layers of silty sandstone with some sandstone/shale

interlayers embedded. Route T-18 runs through five major faults. Dip slope outcrops are often

found on the up-slope along the route.

LANDSLIDE DATABASE

Rainfall-induced landslides along the mileage between 21.5km and 83.5km of Route T-18 are

documented. In total, 55 failed unprotected slopes along T-18 were extracted from the route

maintenance records. Among them, 12 slopes failed during Typhoon Herb, 18 during Toraji, 9

6
during Nari, and 16 during Mindulle due to heavy rainfalls. To match the number of failed

slopes, 54 not-failed unprotected slopes were carefully chosen for the subsequent discriminant

analysis. The not-failed slopes were checked in the field to ensure that they survived all four ty-

phoons and that no evidence of previous slope failure was visible. The locations of failed and

not-failed slopes in the database are shown in Figure 3. Note that the not-failed slopes in the da-

tabase are intentionally chosen to be roughly uniformly distributed over the Route T-18 section.

Also notice that the term “failed slope” in this paper denotes a slope that was destabilized by a

past rainfall event, and the term “not-failed slope” denotes a slope that was not destabilized by

past rainfall events. There are slopes that failed due to other factors, e.g. earthquake, but those

slopes are not the focus of this paper.

The data format is as follows: D = {(xi,ti): i = 1,…,p}, where xi  R10 contains the values

of the 10 landslide features of the i-th slope in the database; ti = 1 if that slope failed, otherwise ti

= 0; p is the total number of slopes in the database. In total, 10 basic features are carefully chosen

based on engineering judgments to reflect the status of a slope: these features are believed to

have significant influence on the stability of a slope. The landslide features are obtained from the

following sources:

(a) Digital maps and digital terrain model (DTM) information

Digital maps with various content and scales are easily available through com-

mercial organizations. In Taiwan, 1/5000 scale maps are available in populated areas and

only 1/10000 scale maps are available in mountainous areas. The DTM (Digital Terrain

Model) employed in this study has a resolution of 40m×40m. It provides digital horizon-

tal and vertical coordinates of the terrain.

(b) Field reconnaissance

7
 Although the method adopted here is to utilize the available geographic infor-

mation and to minimize the field work, the field reconnaissance is still inevitable. For

example, the items of block size and rock volume percentage on slope surface (Note:

these two landslide features will be explained later) are unable to be determined from

aerial photos nor satellite images.

Basic features

Ten basic features are categorized into natural features and man-made features. Among them,

eight are natural features, and two are man-made features (Table 1). The natural features cover

the aspects of topography, geologic conditions, vegetation cover, catchment area of surface wa-

ter, and peak ground acceleration (PGA) induced by a major earthquake. The man-made features

quantify the impacts induced by road construction. Tables 2 and 3 list the data in the database.

The detailed definition of each feature is described in the following section.

Natural features

Topography

Slope direction is defined to be 0o when slope is facing north and increases clockwise. Slope an-

gle is the angle between slope face and road, and slope height is the height of a slope. The sig-

nificance of the direction, angle and height of a slope is demonstrated in Figure 4. Road curva-

ture is the reciprocal of the radius of the route at a slope location (the curvature is defined to be

positive if the slope is in the convex side of the road and negative if the slope is in the concave

side; see Figure 4). Slope direction, slope angle and height are determined from field reconnais-

sance, while road curvatures are determined with the aid of ArcGIS software.

Geology

8
Depending on the age of the outcrop strata, the type of geologic strata at a slope location is quan-

tified into five levels: integers 1 to 5. Integer 1 is given to Zuolan stratum outcrops (the young-

est) and integer 5 is given to Nanzuan stratum outcrops (the oldest).

Vegetative cover

Vegetative cover is divided into two features: area percentage of vegetative cover and thickness

of canopy cover. The former stands for the percentage of vegetative cover on slope faces, while

the latter is the thickness of tree crown. These are indices of vegetative protection against rainfall

scouring on slope surface. Both features need to be obtained from visual inspection during field

reconnaissance.

Water condition

Surface runoff water and groundwater are two major issues for slope stability, and they are quan-

tified in terms of the size of catchment area. The size of catchment area is estimated by DTM:

First interpolate DTM into a smooth 3-D terrain, where the target slope is located. The catchment

area of the slope is sketched manually with the visual aid from the terrain and elevation contours.

Seismicity

Much evidence has shown that since the Chi-Chi earthquake (M = 7.3) which occurred in 1999,

the number of landslides has increased. The peak ground acceleration (PGA) at the slope location

during the Chi-Chi earthquake is an indication of the effect of this earthquake. In total, twenty

one seismic monitoring stations were set up by the Central Weather Bureau of Taiwan around the

Route T-18. The PGA at a particular slope location is then interpolated from the PGA recorded at

the stations during the Chi-Chi earthquake. Exceptions are the slopes that failed in typhoon Herb,

which arrived prior to the Chi-Chi earthquake. For simplicity, the PGA for the slopes that failed

in typhoon Herb are taken to be zeros since the impacts of earlier earthquakes is quite negligible

9
compared to that of the Chi-Chi earthquake.

Man-made features

All slopes considered in this study are subjected to man-made hazards: road construction. Two

features are considered: excavation height at toe and change of slope grade due to toe cutting.

The change of slope grade is defined as the difference in slope angle before and after toe cutting

(see also Figure 4). If landslides occur above the cut slope, then the change of slope grade is

taken as zero. Both features are obtained from field reconnaissance.

LANDSLIDE ANALYSIS

The purpose of the landslide analysis is to predict the locations and occurrence times of land-

slides due to heavy rainfall in future typhoons. The analysis is accomplished by the following

two-step approach:

1. In the first step, called the landslide location analysis, the locations of “rainfall susceptible

slopes” are identified from their 10 basic features in Table 1, not including the rainfall

amount. This identification is conceptually simple: calculate the average degree of similarity

for the 10 basic features between the under-investigation slope and the 55 failed slopes in the

database, then compare it with the average similarity between the slope and the 54 non-failed

slopes. The under-investigation slopes that are more similar to the 55 failed slopes are re-

garded as rainfall susceptible slopes. Such an analysis was conducted for many slopes along

the route to identify rainfall susceptible slopes.

2. In the second step, called the landslide occurrence time analysis, the analysis focuses only on

the rainfall susceptible slopes identified in the first step and only on the rainfall condition. In

particular, the functional relationship between landslide potential and rainfall amount is built

10
 for the identified rainfall susceptible slopes. With this functional relationship and the rainfall

amount predicted by weather forecast, it is then possible to predict the occurrence time of

landslides of these rainfall susceptible slopes.

The rationale behind the two-step approach is as follows. It is found that there are quite a lot

of slopes that never failed during past typhoons regardless the rainfall amount brought by the ty-

phoons, e.g.: slopes with less slope height and angle, slopes with very shallow bedrock, slopes

with excellent vegetation, etc. The 54 non-failed slopes in the database are exactly of this type.

Note that these safe slopes stayed intact not because they were subjected to less rainfall. In fact,

some of these safe slopes experienced significant rainfall amount judging from the historical

rainfall data from the nearby rainfall stations. Therefore, it is assumed that these safe slopes dif-

fer from the failed ones mainly in their 10 basic features rather than in the received rainfall

amount. This is the reason why in the first-step analysis only the 10 basic features are taken into

consideration to compute the average degree of similarity and why the rainfall amount is ex-

cluded from the first-step analysis.

The existence of these safe slopes makes it difficult to conduct the landslide occurrence time

analysis. This is because these slopes will not fail regardless the amount of rainfall received.

Therefore, the key idea in this paper is to screen out the rainfall susceptible slopes in the

first-step analysis, and the identified safe slopes will not go further into the second-step analysis,

i.e. the occurrence time analysis. This treatment eliminates the afore-mentioned difficulty for the

occurrence time analysis. Once the rainfall susceptible slopes are identified in the first step, it is

assumed in the second-step analysis that their landslide occurrence times are mainly governed by

the rainfall conditions, including the rainfall amount and the size of catchment area, but not on

the basic features any longer.

11
First step – landslide location analysis

As mentioned previously, it is desirable to screen out rainfall susceptible slopes in this analysis.

A methodology is proposed to find the locations of such rainfall susceptible slopes along the

mountain road. Predicting landslide occurrence times will the subject of a later section.

A single index is used to quantify the average similarity of an under-investigation slope to

the failed slopes in the database: P (t  1| x, D) , i.e. the probability that t = 1 given x and D,

where x contains the ten basic features of the slope of interest; t denotes the membership of the

slope, i.e. t = 1 if it belongs to the failure group and t = 0 if it belongs to the non-failure group; D

is the data in our database. Later the quantity P (t  1| x, D) will be directly called “predicted

landslide potential”. A slope is called a “rainfall susceptible” slope if its predicted landslide po-

tential is greater than 50%, otherwise it is called a “safe” slope. Note that a “failed slope” is a

slope in the database that actual fails, but a “rainfall susceptible slope” is a slope under investiga-

tion (usually not in the database) whose predicted landslide potential is high.

In this paper, the Gaussian Process analysis (MacKay 1998; William and Rasmussen 1996;

Neal 1997; Gibbs 1997) is implemented to estimate the landslide potential P (t  1| x, D) . The

discriminant function analysis (Fisher 1936), which has been implemented by many previous re-

searchers to predict landslide potentials, is also implemented to compare with the Gaussian Pro-

cess analysis. For the discriminant function analysis, the commercially available program SPSS

(SPSS 1999) is employed. Due to the limitation of paper length, the details for the discriminant

function analysis are not presented here. Only the details for the Gaussian Process analysis are

described in the following section.

One drawback of the discriminant function analysis is that it has to specify the functional

form of the discriminant function, or equivalently, the functional form of the separating boundary

12
of the two membership groups in the feature space. In the absence of physical basis for the speci-

fication, a usual choice is to take the weighted sum of the features (weights are unknown and to

be determined) as the discriminant function. However, it is highly non-trivial to appropriately

specify the functional form of the discriminant function if the problem at hand is highly compli-

cated and not well understood such as the landslide problem. Inappropriate specification may

lead to a large bias.

For complicated problems, it is desirable to adopt an approach that is flexible in the sense

that the discriminant function can be potentially arbitrary. The discriminant function analysis is

not qualified for this purpose. Nonetheless, the neural network qualifies for this purpose as long

as the number of nodes in the network is sufficiently large. However, the required computational

cost for neural network analysis is quite high because when implementing neural network analy-

sis, a non-trivial high-dimensional non-convex optimization problem must be solved.

On the other hand, the Gaussian Process analysis has been recently developed in the field of

artificial intelligence for non-parametric regression and classification. It preserves the advantage

of the neural network: the discriminant function can be potentially arbitrary, but its required

computational cost is much less than that required by neural network analysis. In fact, Neal

(1996) showed that Gaussian Process models are equivalent to neural networks with an infinite

number of nodes whose weights are Gaussian random variables. This result is essential: in order

to implement infinite-node neural networks, an infinite number of parameters are needed, so the

computational cost can be extremely high, but the same task can be achieved by using Gaussian

Process models with only a few parameters.

Brief introduction to the Gaussian Process model for discriminant functions


A Gaussian process in Rn is a stochastic process Y ( x) : x  R n  such that any finite combina-

13
tion of Y ( x1 ), Y ( x2 ),..., Y ( xm ) is jointly Gaussian. Similar to multivariate Gaussian random

variables, a Gaussian Process model is fully defined by the mean, variance, and covariance of the

process.

Suppose the discriminant function Y is known (although in fact it is not), the landslide po-

tential of a slope whose features are given as x is modelled as:

1
P (t  1| Y (.), x)   sig Y ( x)  (1)
1  e Y ( x )

where t indicates the group membership, i.e. t = 1 if the slope belongs to the failure group and =

0 otherwise; the discriminant function Y is, in fact, uncertain and is modeled as a Gaussian Pro-

cess; sig   denotes the sigmoid function. The uncertain discriminant function Y is to be de-

termined, and our goal is to estimate this function by using the past data D = {(xi,ti): i = 1,…,p}.

Instead of assuming the functional form of the Y function (as done in the discriminant func-

tion analysis), in the Gaussian Process analysis, it is only necessary to assume that the Y function

is “smooth” in a certain sense. Note that this smoothness constraint is desirable for sensible pre-

diction. In the context of Gaussian Processes, this smoothness requirement can be enforced by

letting Y ( xi ) and Y ( x j ) be highly correlated when xi and x j are “similar”. Physically, this

smoothness constraint means that if the features of two slopes are similar, then the landslide po-

tentials will be also similar, exemplified by the following covariance model:

 
2
n x( k )  x( k )
1

i j

Cov Y ( xi ), Y ( x j ) H   1  e rk2

2 k 1
  2  C ( xi , x j ) (2)

where xi( k ) denotes the k-th feature in xi ; the exponent is the negative one-half of the square

of the weighted Euclidean distance between xi and xj; Cov denotes covariance. One can verify

that if the xi-xj distance is small (i.e. xi and xj are similar), the covariance (or correlation) between

14
Y(xi) and Y(xj) will be large, and vice versa; 1 , 2 , r1 , r2 ,..., rn   H are the uncertain model pa-

rameters, called the hyper-parameters.

It is worthwhile to discuss the significance of the hyper-parameters. Among them, 1 +  2

gives the variance of Y(x):  2 gives the baseline variance of Y(x) (the variance of the uncertain

mean value of Y(x)) and 1 gives the variance of Y(x) besides the baseline; rk governs the im-

portance of the k-th feature: if rk is small, the k-th feature is influential, and vice versa.

Estimation of landslide potential using the Gaussian Process model

The goal of the first-step analysis is to estimate the landslide potential P(t  1| x, D) . If the dis-

criminant function Y is known, one can see that

1
P(t  1| x, D, Y (.))  P(t  1| x, Y (.))  sig (Y ( x))  (3)
1  e Y ( x )

i.e. the estimation of P (t  1| x, D) is trivial if Y is given. Therefore, the estimation of Y func-

tion is essential for the estimation of P(t  1| x, D) . Unfortunately, the estimation of the Y func-

tion based on D is highly non-trivial. This is because the Y function is only indirectly observed in

the data D through (3): if in the data tj = 1, one can see Y(xj) is more likely to be positive, and if tj

= 0, Y(xj) is more likely to be negative.

Let us now start from an easier position: in the case when the Y function is directly ob-

served at {xi: i = 1,…,p} and when the hyper-parameters H are also given, the estimation of the Y

function is simple. Let us denote ai as the observed value of Y(xi). For simplicity of notation, we

denote the vector [a1 a2 … ap]T by a1:p, [Y(x1) Y(x2) … Y(xp)]T by Y1:p, and [t1 t2 … tp]T by t1:p,

respectively. Due to the Gaussian inference theorem, the updated (posterior) probability density

function (PDF) of Y(x) given a1:p and H is also Gaussian with the following mean and variance:

15
     a1: p    H , a1: p 
1
E Y ( x) | H , a1: p   Cov Y ( x), Y1: p H Var Y1: p H

Var Y ( x) | H , a1: p  (4)

         H , a1: p 
1
 Var Y ( x) | H   Cov Y ( x), Y1: p H Var Y1: p H
T
Cov Y ( x), Y1: p H

where

Var Y ( x) | H   C ( x, x)  1   2

 
Cov Y ( x), Y1: p H  C ( x, x1 )  C ( x, x p )  (5)
 1   2 C ( x1 , x2 )  C ( x1 , x p ) 
 1   2  

Var Y1: p 
H 


 C ( x p 1 , x p ) 
 
 symmetry 1   2 

In reality, the Y function is not directly observed; instead, only {ti: i = 1,…,p} are observed.

Moreover, the hyper-parameters H are uncertain as well. The goal of the Gaussian Process analy-

sis is to estimate P(t  1| x, D) . According to the Theorem of Total Probability (Ang and Tang

1975):

P(t  1| x, D)   P(t  1| x, H , D, a1: p , Y ( x)) f (Y ( x) | H , D, a1: p ) f (a1: p , H | D) dY ( x)da1: p dH (6)

where f ( | ) denotes conditional PDF. Because

P (t  1| x, H , D, a1: p , Y ( x))  P (t  1| H , Y ( x))  sig (Y ( x))

(7)
f (Y ( x) | H , D, a1: p )  f (Y ( x) | H , a1: p )

(6) can be written as

P (t  1| x, D )   sig (Y ( x )) f (Y ( x ) | H , a1: p ) f ( a1: p , H | D ) dY ( x ) da1: p dH (8)

Note that f (Y ( x) | H , a1: p ) is a Gaussian PDF whose mean and variance are specified in (4).

According to MacKay (1992), the integral of the product of a sigmoid function and a Gaussian

PDF can be approximated as

16
 sig (Y ( x)) f (Y ( x) | H , a 1: p 
)dY ( x)  sig   ( H , a1: p ) 1    ( H , a1: p ) 8  (9)

so (8) becomes


P (t  1| x, D )   sig   ( H , a1: p ) 
1    ( H , a1: p ) 8  f ( H , a1: p | D ) da1: p dH (10)

Therefore, the estimation of P(t  1| x, D) involves the evaluation of the high-dimensional in-

tegral in (10), which is a highly non-trivial task.

In this paper, the integral in (10) is estimated through the Law of Large Number (Feller

1968):

1
 sig    ( Hˆ 
N
P (t  1| x, D )  (i )
, aˆ1:(ip) ) 1    ( Hˆ ( i ) , aˆ1:( ip) ) 8 (11)
N i 1

where aˆ1:(ip) and Hˆ ( i ) are the i-th sample pair drawn from f (a1: p , H | D) . In the appendix, a

stochastic simulation procedure based on the Gibbs sampler (Geman and Geman 1984) and hy-

brid Monte Carlo (Duane et al. 1987) is presented to efficiently draw samples from

f (a1: p , H | D ) . The estimated integral is exactly the estimated value of the landslide potential of

the slope of interest.

What follows summarizes the algorithm of estimating landslide potential P (t  1| x, D ) :

(a) Follow the procedures described in the appendix to draw samples

 aˆ
(i )
1: p 
, Hˆ (i ) : i  Tb ,..., T  that are distributed as f (a1: p , H | D) , where Tb and T denote

the end of burning period (see the appendix) and the total sample duration.

(b) For each sample obtained in (a), compute sig   ( Hˆ (i ) , aˆ1:(ip) )  

1    ( Hˆ (i ) , aˆ1:(ip) ) 8 ,

where  ( Hˆ ( i ) , aˆ1:(ip) ) and  ( Hˆ (i ) , aˆ1:(ip) ) are defined in (4).

(c) The average value of the sig   ( Hˆ (i ) , aˆ1:(ip) )  1    ( Hˆ (i ) , aˆ1:(ip) ) 8  values obtained in

17
 (b) is exactly the estimated value of P (t  1| x, D ) .

Results of landslide location analysis

Figure 5 shows the landslide potential map predicted by the analysis for the entire Route T-18.

This map is obtained by interpolating the predicted landslide potentials of many selected slopes

along the route. The dark regions are high landslide potential segments, i.e.: locations of rainfall

susceptible slopes, and light regions are low potential segments, i.e.: locations of safe slopes.

This figure is beneficial for the route maintenance team to predict the locations of potential land-

slides along the route in future typhoons. Note that if a slope is identified as a safe slope, it is not

necessary to conduct the second-step analysis on it because the slope is predicted to be safe re-

gardless the received rainfall amount. Only the identified rainfall susceptible slopes need to be

further analyzed.

Comparison of prediction errors

It is of interest to know how accurate the aforementioned analysis is. Common practice of ex-

amining the performance is to quantify the so-called training errors, which are described as fol-

lows: Once the adopted model is trained by the data D, the trained model is used to find the

landslide potential for each slope in the database, i.e. compute P (ti  1| xi , D) for i = 1, …, 109

in this study. If the landslide potential P (ti  1| xi , D) is larger than 50%, the i-th slope is pre-

dicted as a rainfall susceptible slope, otherwise, it is predicted as a safe slope. If the prediction

shows a non-failed slope is a rainfall susceptible slope, i.e.: the probability that it belongs to the

failure group is greater than 50%, the prediction is inconsistent with the non-failure reality. Sim-

ilarly, if the prediction shows a failed slope is a safe slope, the prediction is also inconsistent with

the reality of failure. If this consistency check is done for all slopes in the training database, the

ratio of inconsistency is then the training error. Unfortunately, this calculation procedure of error

18
is not fair. Because the i-th slope is already within the training data when its landslide potential is

to be predicted, the resulting training error cannot effectively reflect the actual prediction errors

on unseen slopes.

For fair calculation of prediction errors, the so-called leave-one-out (LOO) prediction errors

of the adopted model are adopted here. The LOO prediction error is an unbiased estimate of pre-

diction error on unseen slopes of the trained model. The basic idea of LOO prediction error is to

mimic the prediction process by removing one data point out of the training dataset and use the

removed data point for prediction testing. The procedure of computing LOO prediction error is

as follows: Remove the i-th data point {xi,ti} from the dataset, and call the remaining database

D~i . Estimate P (ti  1| xi , D~i ) . If P (ti  1| xi , D~i ) is larger than 50%, the i-th slope is predicted

as a rainfall susceptible slope, otherwise, it is predicted as a safe slope. Compare the prediction

with the actual status of the i-th slope. Do so for i = 1, …, 109, and the ratio of inconsistency is

exactly the LOO prediction error.

Table 4 shows the traditional training error ratios and the LOO prediction error ratios of the

Gaussian Process analysis and the discriminant function analysis. Note that both training errors

are quite small, but these are not realistic estimates for the actual prediction errors on unseen

slopes. The LOO prediction errors, which more realistically reflect the actual prediction error

rates, are always larger than the training error rates. It is also clear that the Gaussian Process

analysis results in a smaller LOO prediction error ratio than the discriminant function analysis,

indicating that the performance of the former is superior.

It is also worth mentioning that the landslide potential predicted by the discriminant func-

tion analysis is found to be either very close to 0 or very close to 1, implying that the associated

uncertainty is small. This contradicts the reality of the problem at hand: there exist many uncer-

19
tainties so it is expected that the estimated landslide potential should not be close 1 or 0. Never-

theless, the landslide potentials calculated by the Gaussian Process analysis do not exhibit this

undesirable behavior.

Relative importance among features

As discussed in a previous section, the rk of Gaussian Process model governs the relative im-

portance of the k-th feature, hence the mean values of the rk samples (obtained from the Gibbs

sampler and hybrid Monte Carlo) quantify the relative importance of the feature: the smaller the

rk mean value, the more important the k-th feature. Table 5 ranks the relative importance of the

10 features, which shows that slope height, change of slope grade, height of toe cutting and road

curvature are the dominant features, among them are the two man-made features. This result im-

plies that the slope stability along mountain roads is noticeably affected by road construction.

It is of interest to verify if the removal of unimportant features will seriously degrade the

performance of the proposed methodology or not. According to the analysis, the LOO prediction

error ratio of the Route T-18 data is only 1.83% (2 false LOO predictions out of 109) when only

the most important five features are taken. It is clear that the performance of the Gaussian Pro-

cess analysis with only five features is exceptional, implying that five features may be sufficient

for accurate predictions. Even so, the number of ten features is not reduced and still used in this

Gaussian Process analysis.

Landslide occurrence time analysis

The goal of the landslide location analysis is to identify the locations of the rainfall susceptible

slopes, i.e. the slopes with high predicted landslide potentials. Besides predicting potential land-

slide locations, it is desirable to additionally predict during which typhoon the rainfall suscepti-

ble slopes will fail. In Taiwan, landslides are mostly triggered by heavy rainfalls during typhoons.

20
As a consequence, we propose a second stage of analysis to predict the occurrence times of land-

slides for the rainfall susceptible slopes. In the occurrence time analysis, the size of catchment

area and “effective” rainfall amount, which will be explained in detail later, are the only two fea-

tures. This is based on the assumption that once the rainfall susceptible slopes are identified in

the first step, their landslide occurrence times are mainly governed by the rainfall amount and the

size of catchment area, but not on the basic features any longer.

The goal here is to develop the relationship between landslide potential and the two features

for rainfall susceptible slopes given the past landslide and rainfall data. This relationship is called

the rainfall fragility graph. In the case where the future rainfall can be effectively predicted a

priori, e.g.: through weather forecast, it is then possible to predict the occurrence times of land-

slides with the rainfall fragility graph.

Database

The database employed here is based on the database for the landslide location analysis. Howev-

er, there are many changes in the database for the occurrence time analysis. Firstly, most features,

except the size of catchment area, are removed from the database, but one new feature, i.e.: ef-

fective rainfall amount, is added into the database so that the total number of features is now two.

In fact, these two features are related to the amount of water supply for a slope. Secondly, the 12

failed slope data points for Typhoon Herb (year 1996) are removed from the database since the

landslide behaviours of the slopes seems to change significantly after the Chi-Chi earthquake

(year 1999). Finally, all not-failed slope data in the landslide location analysis is removed. This is

because, as mentioned previously, these slopes did not fail regardless of the received rainfall

amount, i.e.: they are the “safe slopes” in the database. Therefore, these slopes need to be re-

moved from the database for the second-step analysis: they cannot be taken as the not-failed

21
slopes of the new database.

In conclusion, only the catchment areas and effective rainfall amount of the failed slopes are

kept in the new database. The effective rainfall amount for a failed slope is taken as the effective

rainfall amount received by the slope during its landslide-causing typhoon. An issue naturally

arises: not-failed data points are absent in the new database, but they are necessary for the analy-

sis. In principle, the necessary not-failed data points here are quite different from the not-failed

data points in the landslide location analysis database: the new database needs slopes belonging

to the failure group but had “not yet” failed because the received rainfall amount is not enough.

Let us call these slope data points as “not-yet-failed slopes”. These data points can be obtained

easily: before the landslide-causing typhoon of a failed slope, the slope did not yet fail, i.e.: each

failed slope in the database can be used to create not-yet-failed data points. This is achieved by

taking a failed slope in the database and replacing its effective rainfall amount by the effective

rainfall amount during each of the typhoons prior to the landslide-causing typhoon. The total

numbers of failed slopes and not-yet-failed slopes in the new database are 43 and 33, respectively.

These data points are shown in Figure 7.

Effective rainfall amount

The actual hourly rainfall data were obtained from 34 rainfall stations located in the mountainous

region of central Taiwan (see Figure 3). However, most of the rainfall stations are not adjacent to

Route T-18. Therefore, the hourly rainfall amount history (in mm) at a slope is determined by

interpolation with the inverse-distance-weighting (IDW) method built in the ArcGIS program

(ArcGIS 2001). Once the hourly rainfall amount history at a slope is obtained, the following

procedure is employed to quantify the effective rainfall amount Reff at the slope caused by a

typhoon: Firstly, the 72-hour average rainfall is calculated at the slope for that typhoon. As an

22
example, let Figure 6 be the interpolated hourly rainfall time history at the slope, and let the time

instant of interest be the end of the typhoon (0 hour in the figure), i.e. the interest now is to find

the amount of water supply to the slope at the end of the typhoon. The 72-hour average rainfall

A72 is defined as the area under the rainfall history between -72 hour and 0 hour in Figure 6(a)

divided by 72 hours. Secondly, a 12-hour moving window is employed to calculate the 12-hour

moving average rainfall in the -72 to 0 hour interval. As a result, a curve of 12-hour moving av-

erage rainfall can be obtained (see the dotted curve in Figure 6(b)). The maximum 12-hour mov-

ing average rainfall M 12 is defined as the maximum height of this dotted curve. Note that A72

is used to quantify the accumulated rainfall in the past 72 hours, while M 12 is used to quantify

the rainfall intensity in the past 72 hours. The effective rainfall amount Reff is defined as the

average of A72 and M 12 :

1 
Reff   A72  M 12  (12)
2 

The above equation takes into account the accumulative effect of the entire rainfall event and al-

so the short-term impact of the rainfall intensity.

Results of occurrence time analysis

Although the goal here is different from that in the landslide location analysis, the same Gaussian

Process analysis can be taken to compute the landslide occurrence potential P(t  1| x, D) ,

where x contains only two features (catchment area and Reff) of the under-investigation rainfall

susceptible slope; t denotes the status of the slope, i.e. t = 1 if it fails and t = 0 if it does not fail;

D is the new database.

Figure 7 shows the results of the Gaussian Process analysis and is called the rainfall fragili-

ty graph. In the figure, the crosses “+” indicate the failed slopes in the new database (43 data

23
points), while the circles “o” are the not-yet-failed slopes (33 data points). Note that in the low-

er-left region, i.e. small rainfall and small catchment area, most rainfall susceptible slopes did not

yet fail, while most rainfall susceptible slopes failed in the upper-right region, i.e. large rainfall

and large catchment area. This observation agrees with intuition. The contour values indicate the

value of the predicted landslide potential P(t  1| x, D) by the Gaussian Process analysis based

on the new database. This rainfall fragility graph can then be used to predict landslide potential

of a rainfall susceptible slope due to a future typhoon.

Prediction errors

Note that in Figure 7, the mixture of the failed and not-yet-failed data points is evident, indicat-

ing that the discrimination between the failed and not-yet-failed slopes is difficult. There are

several possibilities for this difficulty: (a) the effective rainfall and catchment area are not the

only controlling features for discriminating the failed cases from not-yet-failed cases, i.e. there

are some missing essential features; (b) Reff is poor in quantifying the rainfall amount; (c) the

IDW method of interpolating the rainfall needs improvement; (d) the union of the above. As a

result, the LOO prediction error for the landslide occurrence time analysis seems higher (see Ta-

ble 6) than that for the landslide location analysis (see Table 4), indicating that landslide occur-

rence time analysis is more challenging than the location analysis.

Implementing rainfall fragility graph

The rainfall fragility graph can be used to estimate the real-time landslide potential of a rainfall

susceptible slope, as demonstrated with the following example. Let us consider the rainfall his-

tory shown in Figure 8 for a rainfall susceptible slope whose catchment area is 27332 m2. Since

the Reff index depends on the past 72-hour rainfall record, the rainfall history can be easily con-

verted in real time to the Reff history, shown in Figure 8. Moreover, at each time instant, one can

24
compute the landslide potential given the Reff information and the catchment area size with the

aid of the rainfall fragility graph in Figure 7. Therefore, the landslide potential can be determined

as shown in Figure 8.

If a rainfall forecast is available, it is also possible to implement the rainfall fragility graph

to predict landslide potential of a rainfall susceptible slope. However, the accuracy of the land-

slide prediction will heavily rely on the accuracy of the rainfall forecast.

CONCLUSION

Evaluating rainfall-induced landslide potentials for slopes along mountain roads is a complicated

matter due to large uncertainties. In this paper, a full probabilistic analysis based on the Gaussian

Process model is proposed to predict rainfall-induced landslide locations and occurrence times

along a mountain road in Taiwan. The landslide database for the slopes along the Alishan moun-

tain road, which contains the features of 55 failed and 54 not-failed slope cases, is adopted to

demonstrate the analysis. The analysis results show that the prediction error is around 5.5% for

the landslide location analysis. The performance of the Gaussian Process analysis is superior to

that of the discriminant function analysis, which is easier to operate than the former but less

flexible. The relative importance of the chosen features is quantified. For the landslide occur-

rence time analysis, a rainfall fragility graph is obtained. Scattered data in the graph indicates

that the uncertainties are significant.

The research results given in this paper include the landslide potential map and the rainfall

fragility graph. The former quantifies the rainfall-induced landslide potentials of the slopes along

the route, so it is valuable for predicting the locations of future rainfall-induced landslides, while

the latter relates the landslide potential with the amount of rainfall and size of catchment area of

25
a rainfall susceptible slope, hence it is valuable for predicting the occurrence times of future

landslides. Note that the implementation of the latter requires accurate prediction of the rainfall

amount, which is challenging research by itself.

Finally, it should be pointed out that the rainfall fragility graph shown in Figure 7 is devel-

oped based on the rainfall data and the nature/man-made features of slopes gathered from Route

T-18 of Taiwan. It is more or less tailor made for Route T-18. Modifications may be needed when

applied to other mountain roads in other locations.

ACKNOWLEDGEMENTS

The authors wish to thank the Second and the Fifth Route Maintenance Divisions of the Taiwan

Highway Bureau for providing the case history records and all the necessary help to carry out the

field observations along Route T-18. The financial supports provided by the Harbor and Marine

Technology Center (Project No. MOTC-IOT-95-H1DB004) and the National Science Council of

the Republic of China (Project No. 94-2625-Z-011-002) are also appreciated.

REFERENCES

ANG A.H.S. and TANG W.H. Probability concepts in engineering planning and design: Volume

I – basic principles, John Wiley and Sons, New York, 1975.

CARRERA A., CARDINALI M., DETTI R., GUZZETTI F., PASQUI V. and REICHENBACH P.

GIS techniques and statistical models in evaluating landslide hazard. Earth Surface Pro-

cesses and Landforms, 1991, 16, 427-445.

CARRERA A., CARDINALI M., GUZZETTI F. and REICHENBACH P. GIS technology in

mapping landslide hazard. in Geographical Information Systems in Assessing Natural

26
 Hazards, 1995.

CHUNG C. F. and FABBRI A. G., Probabilistic prediction models for landslide hazard mapping,

Photogrammetric Engng and Remote Sensing, 1999, 65, No. 12, 1389-1399.

DUANE S., KENNEDY A. D., PENDLETON B.J. and ROWETH D. Hybrid Monte Carlo.

Physics Letters B, 195, 1987, 216-222.

ARCGIS. Version 8.2, ESRI, Redlands, California, USA, 2001.

FELLER W. An introduction to probability theory and its applications: volume I, John Wiley and

Sons, New York, 1968.

FISHER R.A. The use of multiple measurements in taxonomic problems. Ann. Eugen., 1936, 7,

179-188.

FURBISH D.J. Debris slides related to logging of streamside hillslopes in northwestern Califor-

nia, Humboldt State University, Arcata, California, U.S, M.S. Thesis, 1981.

GEMAN S. and GEMAN D. Stochastic relaxation, Gibbs distribution and the Bayesian restora-

tion of images. IEEE Trans. Pattern Anal. Machine Intell, 1984, 6, 721-741.

GIBBS M. N. Bayesian Gaussian processes for regression and classification, University of

Cambridge, PhD thesis, 1997.

GORSEVSKI P.V., GESSLER P.E. and FOLTZ R.B. Spatial prediction of landslide hazard using

logistic regression and GIS. Proc. of 4th International Conference on Integrating GIS and

Env. Modeling, Banff, Alberta, Canada, 2000.

KOJAN E., FOGGIN III G.T. and RICE R.M. Prediction and analysis of debris slide incidence

by photogrammetry. Proc. of 24th International Geology Congress, 1972, Section 13,

124-131.

27
LIAO H.- J., LIN Y.- C. and LIAO J. T. Evaluation of slope stability with discriminant analysis.

Proc. of 14th Southeast Asian Geotech Conference, Hong Kong, Balkema, 2001,

1131-1134.

MACKAY D.J.C. Introduction to Gaussian processes. in Neural Networks and Machine Learn-

ing, 1998, vol. 168 of NATO Asi Series. Series F, Springer Verlag.

MACKAY D.J.C. A practical Bayesian framework for backpropagation networks. Neural Com-

putation, 1992, 4, No. 3, 448-472.

MARK R.K. and ELLEN S.D. Statistical and simulation models for mapping debris-flow hazard.

Geographical Information Systems in Assessing Natural Hazards, 1995, 93-106.

NEAL R. M. Monte Carlo implementation of Gaussian process models for Bayesian regression

and classification. Dept. of Computer Science, University of Toronto, 1997, Technical

Report CRG-TR-97-2.

NEAL R.M. Bayesian learning for neural networks. Lecture Notes in Statistics 118, Springer,

1996.

PILLSBURY N.H. A system for landslide evaluation on igneous terrain. Colorado State Univer-

sity, Fort Collins, Colorado, U.S, PhD Thesis, 1976.

RICE R.M. and FOGGIN III G.T. Effect of high intensity storms on soil slippage on mountain-

ous watersheds in southern California. Water Resour. Res., 1971, 7, No. 6, 1485-1495.

RICE R.M., PILLSBURY N.H. and SCHMIDT K.W. A risk analysis approach for using discri-

minant functions to manage logging-related landslides on granitic terrain. Forest Science,

1985, 31, No. 3, 772-784.

SANTACANA N., BAEZA B., COROMINAS J., PAZ A.D. and MARTURIA J. A GIS-based

multivariate statistical analysis for shallow landslide susceptibility mapping in La Pobla de

28
 Lillet Area. Natural Hazards, 2003, 30, 281-295.

SPSS. SPSS for Windows, 1999, Chicago: SPSS Inc.

WALTZ J.P. Analysis of selected landslides in Alameda and Contra Costa counties, California.

Bull. Assoc. Eng. Geol, 1971, 8, No. 2, 153-163.

WEN S.-L., LIAO H.-J., LIN Y.-C. and LAN Y.-S. Evaluation study on probability of landslide

in hillside community by discriminant analysis. Architecture Inst. of Republic of China (in

Chinese), J. of Architecture, 2001, 37, 49-63.

WILLIAMS C.K.I. and RASMUSSEN C.E. Gaussian Processes for regression. in Advances in

Neural Information Processing Systems 8, MIT Press., 1996.

APPENDIX: ALGORITHMS FOR DRAWING SAMPLES FROM f (a1: p , H | D)

In the main text, it is mentioned that drawing a1: p and H samples from f (a1: p , H | D) is re-

quired for the estimation of P(t  1| x, D) . In this appendix, the procedure for achieving this

based on the Gibbs sampler and hybrid Monte Carlo method is presented.

Gibbs sampler

The Gibbs sampler is a powerful Markov chain Monte Carlo technique. The idea is to sam-

ple from a special Markov chain whose states are the target variables a1: p and H and whose

stationary distribution is the target PDF f (a1: p , H | D) . The basic principle of the Gibbs sampler

is to divide target variables into groups, and to sample each group conditioning the others. For

our purpose, the target variables a1: p and H are divided into two groups: a1: p and H . The

Gibbs-sampler procedure of sampling a1: p and H is as follows:

(a) Initialize aˆ1:(0)p and Ĥ (0) at any values.

29
 
(b) Sample aˆ1:(ip) ~ f a1: p | Hˆ (i 1) , D  and Hˆ (i ) ~ f  H | aˆ1:(ip) , D 

(c) Repeat step (b) for time steps i = 1,…,T to obtain  aˆ(i )
1: p  
, Hˆ (i ) : i  0, , T . These sam-

ples will be asymptotically distributed as f (a1: p , H | D) if the Markov chain is ergodic.

Since the Markov chain starts from an arbitrarily chosen location aˆ1:(0)p and Ĥ (0) , it re-

quires some time for the Markov chain to become stationary. Samples prior to stationarity are, in

general, not distributed as f (a1: p , H | D) , so they should be discarded. The time period prior to

the stationarity is called the burning period. The burning period is usually identified by visual

inspection of sample time histories. Only the samples after the burning period will be taken for

the estimation of P(t  1| x, D) .

Hybrid Monte Carlo

In the Gibbs sampler algorithm, it is necessary to sample from two PDFs: f a1: p | Hˆ (i 1) , D  
and f  H | aˆ1:(ip) , D  , and this is not a trivial task. The hybrid Monte Carlo technique is employed

here to achieve the sampling, which is another Markov chain Monte Carlo technique that is

highly efficient when the target variables are high dimensional. The use of hybrid Monte Carlo

requires the ability to evaluate the target PDF up to a constant. According to the Bayes’ rule,

 
f a1: p | Hˆ (i 1) f  D | a1: p 
 
f a1: p | Hˆ (i 1) , D   
 f a1: p | Hˆ (i 1) f  D | a1: p 

f D | Hˆ ( i 1)

f  aˆ1:(ip) | H  f  H 
f  H | aˆ , D   f  H | aˆ
(i ) (i )
  f  aˆ1:(ip) | H  f  H  (13)
f  aˆ 
1: p 1: p (i )
1: p

Therefore, the use of the hybrid Monte Carlo technique requires the ability to evaluate

f  a1: p | H  , f  D | a1: p  , and f  H  . Note that f  H  is the prior PDF of H that reflects our

30
prior belief on the possible value of H without the data D. It is chosen by the user, so its evalua-

tion is simple. In this paper, f  H  is chosen to be a constant function to reflect our lacking of

knowledge of H without the data. The mathematical expressions of f  a1: p | H  and f  D | a1: p 

are as follows:

 
1 1
1 1
f  a1: p | H  
 a1:TpVar a1: p H a1: p
2
e
 2    
p/2
det Var a1: p H

1ti
 1   e i 
p a t

f  D | a1: p    
i

 ai   a 
(14)
i 1  1  e  1 e i 

To explain the hybrid Monte Carlo procedure, let us now denote the target PDF by f  x  for

discussion convenience, where x is the target vectoral variable, i.e. a1: p or H , f  x  can be


either f a1: p | Hˆ (i 1) , D  or f  H | aˆ1:(ip) , D  (note that the condition notation has been omitted

for simplification). The PDF f  x  can be evaluated up to a constant, i.e. f  x   a  h  x  ,

where h  x  can be evaluated. The basic idea of the hybrid Monte Carlo method is to add an

auxiliary independent standard Gaussian random variable z to the process, so the new target PDF

is proportional to
n
1
  z 2j
f  x, z   h  x   e
2 j 1
(15)

where n is the dimension of x and z; zj is the j-th component of the z vector. If we are able to

sample from f  x, z  , it is clear that the x part of the samples will be distributed as f  x  .

The main trick for hybrid Monte Carlo is to give h( x) and z some physical meaning:

 log  h( x)  is considered to be the potential energy of a ball with unit mass (x is the location of

31
the ball; think of  log  h( x ) to be the terrain profile of a mountain valley) and z is the velocity

of the ball. So the total energy of the ball is

1 n 2
 log h  x    z j   log f ( x, z )  H  x, z 
2 j 1
(16)

Since there is no friction when the ball rolls on the valley side, the total energy H  x, z  is con-

servative, i.e. f ( x, z ) is constant when the ball rolls according to the following Hamiltonian

equations:

dx j dz j H  x, z   log h  x 
 zj   j  1,..., n (17)
dt dt x j x j

where xj and zj are the j-th components of the x and z vectors. That is to say, the ball follows

equal-probability lines in the (x,z) space (phase space). To let the ball be able to travel throughout

the entire phase space, the velocity z is occasionally resampled from the standard Gaussian PDF

to alter the total energy. By doing so, it can be shown that the (x,z) samples will be asymptotical-

ly distributed as f  x, z  , and the x part of the samples will be asymptotically distributed as the

target PDF f  x  .

The procedure of the hybrid Monte Carlo method is summarized as follows:

(a) Initialize x̂ (0) , ẑ (0) at any specified location

(b) Solve (x,z) trajectory, i.e. x(t) and z(t) time histories, according to Hamiltonian equation

dx j (t ) dz j (t ) H  x(t ), z (t )   log h  x(t ) 

 zj   j  1,..., n (18)
dt dt x j x j

with initial condition x(0)  xˆ (i 1) and z (0)  zˆ (i 1) . Solve the equations for a random-

ized duration, and let xˆ (i ) be the end location of the solution.

(c) Resample zˆ (i ) from the standard Gaussian PDF.

32
 (d) Cycle (b)-(c) for i = 1,…, T to get  xˆ (i )
: i  0,..., T  . These samples will be asymptoti-

cally distributed as f  x  if the Markov chain is ergodic.

In practice, step (b) cannot be solved analytically. To handle this issue, we can adopt the

following “leapfrog” algorithm to solve the Hamiltonian equations approximately:

 t  t h  x  xi x  xt 
zi  t    zi  t  
 2  2 h  x t 
 t 
xi  t  t   xi  t   t  zi  t   (19)
 2 

 t  t h  x  xi x  x t t 
zi  t  t   zi  t   
 2  2 h  x  t  t  

where t is the length of unit time step. When implementing the leapfrog algorithm, an ac-

cept/reject step is required to assure the so-called detailed balance condition required by the

Markov chain Monte Carlo methods: Let xˆ C and zˆ C (the superscript C denotes “candidate”)

be the end location and velocity from the leapfrog algorithm for a randomized duration in step

(b). Accept the candidate, i.e. take xˆ (i )  xˆ C , with probability r, where

  zˆCj 
n
1 2

f  xˆ , zˆ  h  xˆ e
C C C 2 j 1

r  (20)
f  xˆ ( i 1) ( i 1)
   zˆ(ji1) 
n
, zˆ 
1 2

h  xˆ (i 1)   e
2 j 1

If not accepted, repeat the previous sample, i.e. take xˆ (i )  xˆ (i 1) .

33
 Table 1 Basic features and their numerical ranges

Categories Landslide features Unit Ranges

Slope direction ° 0~360
Natural Features

Slope angle ° 0~90

Topography
Slope height m ≧0
Road curvature 1/m -∞~+∞
Geology Strata type - 1~5
Vegetative cover Thickness of canopy cover m ≧0
Water condition Catchment area m2 ≧0
Seismicity Peak ground acceleration gal ≧0
Man-made feature Height of toe cutting m ≧0
- Road construction Change of slope angle ° 0~90

34
 Table 2 Route T-18 database (failed slopes)

Slope Slope Road Thickness of Height of Change of Peak ground

Mileage Slope
# angle height curvature Strata canopy
Catchment
toe cut- slope acceleration
(Km+m) direction (°) type area (m2)
(°) (m) (1/m) cover (m) ting (m) grade (°) (gal)

1 20k+750 330 70 30 -0.048 3 2.5 27332 2 0 201.3

2 21k+100 30 80 30 -0.008 3 1.5 2845 12 20 204.8
3 21k+400 100 40 25 -0.007 3 3.5 15343 2 0 202.2
4 21k+500 40 60 30 -0.006 3 2 3353 5 10 202.2
5 21k+555 30 60 20 -0.004 3 2 2134 5 10 202.2
6 21k+650 120 60 20 0.001 3 2 12396 5 0 0.0
7 21k+900 40 50 20 -0.009 3 2 20118 4 0 0.0
8 22k+050 35 65 25 -0.005 3 1.5 3150 15 15 0.0
9 22k+250 110 60 25 -0.008 3 2.5 9754 3 0 0.0
10 22k+350 175 60 25 -0.026 3 2 11380 7 0 196.1
11 22k+350 175 60 25 -0.026 3 2 11380 7 0 196.1
12 22k+850 0 60 25 0.002 2 3.5 7925 8 0 194.8
13 22k+850 0 60 25 0.002 2 3.5 7925 8 0 194.8
14 22k+950 340 50 20 -0.002 2 3.5 48873 8 0 194.8
15 24k+700 145 60 40 -0.018 2 3.5 4369 2 0 197.4
16 24k+750 175 75 15 0.012 2 3 406 8 15 197.4
17 24k+860 170 80 25 -0.024 3 1.5 1118 12 10 200.9
18 24k+900 165 70 20 0.014 3 1.5 406 12 10 197.4
19 24k+900 165 70 20 0.014 3 1.5 406 12 10 0.0
20 26k+100 130 50 15 -0.008 1 2.5 19102 5 0 202.6
21 26k+800 215 45 30 -0.023 3 1.5 80574 6 0 225.5
22 26k+800 215 45 30 -0.023 3 1.5 80574 6 0 225.5
23 26k+800 215 45 30 -0.023 3 1 80574 6 0 225.5
24 28k+100 150 50 40 -0.016 5 4.5 47552 3 0 203.4
25 28k+100 150 50 40 -0.016 5 2 47552 3 0 203.4
26 31k+200 265 60 18 0.004 5 0.5 11685 8 20 198.2
27 31k+700 275 40 20 -0.012 5 3 132901 4 0 195.8
28 40k+400 155 60 35 -0.004 5 1.5 1422 5 0 191.4
29 42k+700 90 75 25 0.002 5 2 3556 6 0 233.6
30 43k+150 130 55 50 0.003 5 2 15343 3 0 237.5
31 45k+180 95 65 40 -0.002 5 1.5 12599 5 0 269.1
32 45k+200 95 70 45 -0.005 5 1 12599 8 5 0.0
33 45k+200 95 70 45 -0.005 5 1 12599 8 5 269.1
34 45k+950 135 65 20 0.002 5 2 5182 12 20 278.5
35 53k+250 245 55 60 -0.007 5 3 16765 6 0 327.6
36 56k+450 120 60 20 0.011 4 1 3963 7 10 321.5
37 57k+050 140 75 40 -0.032 4 2 183534 7 0 330.8
38 58k+000 140 65 100 -0.017 5 0.5 43966 15 10 336.2
39 58k+400 95 75 60 -0.039 5 1 55695 13 0 0.0
40 58k+400 95 75 60 -0.039 5 1 55695 13 0 341.1
41 58k+500 135 75 25 0.021 5 1 1422 10 25 341.6

35
42 58k+950 160 50 50 -0.012 5 3.5 74797 5 0 0.0
43 59k+500 155 60 30 -0.017 5 2 30380 8 15 351.9
44 63k+400 240 60 25 -0.038 5 1 166635 6 10 380.3
45 65k+000 170 70 55 -0.015 5 0.5 255743 50 10 374.2
46 65k+750 235 70 50 -0.032 5 0.5 16460 35 10 373.2
47 66k+000 190 70 50 -0.003 5 1 1626 30 10 0.0
48 66k+200 215 85 25 0.007 5 1 1016 23 20 372.0
49 66k+700 185 80 23 0.017 5 1 2134 16 20 371.4
50 69k+300 200 65 20 -0.004 5 2 29118 8 20 382.1
51 69k+400 240 40 25 -0.019 5 4 97373 4 0 0.0
52 69k+600 265 50 20 -0.025 5 4 31904 4 0 376.4
53 73k+400 210 60 15 0.003 5 3 4674 6 20 388.0
54 74k+800 0 70 25 -0.023 5 3 112682 9 30 0.0
55 77k+050 185 80 23 0.013 5 0.5 3048 14 20 0.0

Table 3 Route T-18 database (not-failed slopes)

Slope Slope Road Thickness of Height of Change of Peak ground

Mileage Slope direc- Strata Catchment
# angle height curvature canopy toe cut- slope acceleration
(Km+m) tion (°) type area (m2)
(°) (m) (1/m) cover (m) ting (m) grade (°) (gal)

1 21k+300 55 65 10 0.009 3 1 1422 6 15 205.6

2 25k+550 25 65 12 0.010 2 3 1829 8 15 210.0
3 26k+300 190 50 15 0.013 1 1 2134 2 0 202.6
4 30k+750 20 70 12 0.010 5 1 5182 3 0 202.6
5 32k+250 300 55 12 0.010 5 3.5 5690 3 0 192.5
6 32k+550 265 60 15 0.013 5 4 1422 3 0 189.2
7 33k+300 315 55 12 0.010 5 2 1290 3 0 186.0
8 33k+700 30 65 10 0.009 5 3 1475 3 0 181.8
9 33k+850 55 60 15 0.013 5 3 2120 3 0 182.9
10 34k+300 260 70 13 0.011 5 2.5 16036 3 0 184.0
11 35k+100 310 65 8 0.007 5 3 737 4 25 178.0
12 36k+300 310 50 12 0.010 5 3 1659 2 0 173.5
13 38k+400 60 90 12 0.010 5 1.5 922 4 30 178.7
14 38k+450 25 70 12 0.010 5 1 1106 4 20 179.9
15 40k+650 100 60 11 0.010 5 2.5 1016 3 0 195.6
16 40k+800 335 65 15 0.013 5 1 2032 4 15 194.5
17 42k+500 45 70 15 0.013 5 2 2337 5 0 209.4

36
18 42k+600 85 60 17 0.015 5 2.5 5284 3 0 209.4
19 43k+500 345 70 11 0.010 5 0.5 2032 7 15 238.8
20 45k+050 90 85 11 0.010 5 1.5 5995 5 20 269.1
21 47k+100 0 70 15 0.013 5 3 1626 4 0 293.9
22 47k+300 310 50 23 0.020 5 4 3455 2 0 295.4
23 47k+400 290 65 18 0.016 5 4 1524 4 0 295.4
24 49k+000 70 65 25 0.022 5 4 813 4 0 311.7
25 51k+000 105 65 25 0.022 5 1.5 3861 4 0 324.4
26 51k+150 80 60 12 0.010 5 1.5 813 3 0 324.4
27 52k+850 295 70 15 0.013 5 1.5 2337 4 10 332.4
28 53k+050 280 60 17 0.015 5 2 4369 4 0 327.6
29 53k+850 280 85 15 0.013 5 1 4267 9 15 317.2
30 54k+150 250 70 20 0.017 4 1 2337 9 15 313.6
31 59k+750 245 30 18 0.016 5 2.5 2032 3 0 352.2
32 59k+950 230 30 17 0.015 5 2.5 2845 3 0 347.2
33 60k+900 50 50 15 0.013 5 3 3861 4 0 357.4
34 61k+250 130 50 12 0.010 5 0.5 1931 5 0 362.5
35 61k+500 130 60 15 0.013 5 1 12802 5 0 362.5
36 61k+950 260 55 7 0.006 5 2.5 813 4 15 357.5
37 62k+400 125 50 17 0.015 5 3 3658 4 0 371.2
38 63k+600 290 65 6 0.005 5 2.5 1422 4 20 375.4
39 65k+200 200 90 20 0.017 5 1 2337 10 15 373.7
40 66k+550 125 70 13 0.011 5 1 2134 8 15 371.4
41 67k+400 80 70 10 0.009 5 3 1626 2 35 380.5
42 70k+300 290 50 13 0.011 5 1 1219 5 15 372.6
43 71k+400 245 70 7 -0.004 5 1 1727 6 30 379.3
44 71k+600 190 55 15 -0.007 5 1 2134 10 20 384.1
45 71k+700 225 60 6 0.011 5 1 2743 5 20 383.1
46 73k+300 200 60 12 0.012 5 0.5 1727 7 20 383.1
47 73k+700 325 60 6 0.008 5 3.5 1931 4 35 391.9
48 75k+000 335 60 5 0.007 5 4 3455 3 30 389.5
49 76k+000 215 60 9 0.017 5 3 914 8 25 377.4
50 78k+950 215 90 7 -0.008 5 1.5 711 5 30 381.2
51 79k+050 225 60 6 0.015 5 3 2642 4 25 381.2
52 79k+650 255 60 5 -0.015 5 3 38204 4 25 376.9
53 80k+600 265 45 5 0.025 5 3 914 4 25 371.6
54 81k+700 270 45 10 -0.019 5 3 24081 7 15 365.3

37
Table 4 Training errors and LOO prediction errors for the landslide location analysis

Training # of False LOO Predictions LOO Prediction

Methodology
errors (out of 109) Error Ratio
Discriminant Function Analysis 7.3% 13 11.9%
Gaussian Process Analysis 0% 7 6.4%
Note: LOO = leave-one-out analysis method

Table 5 Relative importance of the ten features

Mean Value of
Landslide Feature Ranking
r Samples
Slope height 0.9437 1
Change of slope grade 1.4227 2
Height of toe cutting 1.6656 3
Road curvature 1.9634 4
Catchment area 2.3562 5
Strata type 2.4306 6
Thickness of canopy cover 2.7721 7
Slope angle 2.7765 8
Peak ground acceleration 2.7841 9
Slope direction 3.6612 10

Table 6 LOO prediction errors for the landslide occurrence time analysis

# of False LOO LOO Prediction

Predictions (out of 76) Error Ratio
LOO predic-
27 35.5%
tion errors

38
 Accumulated rainfall (mm)

Figure 1 Accumulative rainfall records at Alishan rainfall station near Route T-18

Figure 2 Typhoon-induced damage types of landslides and case numbers along Route T-18

39
 km

Figure 3 Locations of failed and not-failed slopes along Route T-18 and the nearby rainfall stations

40
 thickness of
canopy cover

slope
change of slope grade
height

toe cutting

dip height of
angle toe cutting

dip
orientation
road slope
curvature(+) angle road
curvature(-)
pe
slo tion
ec
dir
road
N

Figure 4 Significance of the topographical features of the slopes in the database

41
Figure 5 Landslide potential map along Route T-18

42
 72-hour accumulated rainfall/72=A72

Time Instant of
Interest
A72

Figure 6(a) Computation of A72

72 hr

12 hr moving average

Time Instant of
M12 Interest

Figure 6(b) Computation of M12

Figure 6 Computation of 72-hour average rainfall A72 and maximum 12-hour moving average

rainfall M 12

43
 Landslide Potential

Figure 7 Rainfall fragility graph for rainfall susceptible slopes along Route T-18

44
 Landslide Potential

Landslide Potential
Hourly Rainfall (mm)

Figure 8 The rainfall, Reff and landslide potential time histories of a demonstrative rainfall susceptible

slope (catchment area = 27332 m2)

45
 Table

Table 1 Landslide features and their numerical ranges

Table 2 Route T-18 database (failed slopes)

Table 3 Route T-18 database (not-failed slopes)

Table 4 Training errors and LOO prediction errors for the landslide location analysis

Table 5 Relative importance of the ten features

Table 6 LOO prediction errors for the landslide occurrence time analysis

Figure

Figure 1 Accumulative rainfall records at Alishan rainfall station near Route T-18

Figure 2 Typhoon-induced damage types of landslides and case numbers along Route T-18

Figure 3 Locations of failed and not-failed slopes along Route T-18 and the nearby rainfall stations

Figure 4 Significance of the topographical features of the slopes in the database

Figure 5 Landslide potential map along Route T-18

Figure 6 Computation of 72-hour average rainfall A72 and maximum 12-hour moving average

rainfall M 12

Figure 7 Rainfall fragility graph for rainfall susceptible slopes along Route T-18

Figure 8 The rainfall, Reff and landslide potential time histories of a demonstrative rainfall susceptible

slope

46

View publication stats