Вы находитесь на странице: 1из 14

Technical Note

Landslides (2010) 7:325338 Nobutomo Osanai I Takeshi Shimizu I Kazumasa Kuramoto I Shinichi Kojima I
DOI 10.1007/s10346-010-0229-5 Tomoyuki Noro
Received: 30 November 2009
Accepted: 14 June 2010
Published online: 24 July 2010 Japanese early-warning for debris flows and slope
Springer-Verlag 2010
failures using rainfall indices with Radial Basis
Function Network

Abstract Early-warning systems for natural disasters are important probability of occurrences of debris ows and slope failures is
tools for disaster risk reduction and for achieving sustainable high and where there are many houses vulnerable to damage.
development and livelihoods. In 2005, the Japanese government However, it is difcult to install structural measures in all
initiated a new nationwide early-warning system for landslides dangerous valleys because there are some 200,000 dangerous
disasters. The main methodology of the system is to set a criterion valleys and slopes according to MLIT statistics. Researchers and
for occurrences of debris ows and slope failures based on several government ofcials in Japan have come to recognize that it is
rainfall indices (60-min cumulative rainfall and soilwater index) important to provide people with warning information to
in each 5-km grid mesh covering all of Japan. Because many of the evacuate from disasters even where structural measures are
records of mass movements are lacking in scientic precision on installed. MLIT has developed a landslide early-warning system
timing and location, the system applies Radial Basis Function since 1984 to protect people from injury, loss of life, and loss of
Network methods to set the criterion based primarily on rainfall livelihood. Damage from landslides is less now than in the past. In
data recorded as not triggering disasters. Since the end of March the rainy season, there are still about 1,000 landslide disasters
2007, under torrential rainfall conditions, early-warning informa- reported annually. These Japanese statistics indicate that early-
tion has been disseminated as part of weather news using TV, radio, warning systems are important not only in developing countries
and the Internet. Because of the increasing worldwide recognition (UN ISDR 2006) but also in developed countries to reduce
of the importance of early-warning systems for natural disaster damage by small-magnitude and high-frequency landslides. In the
reduction, the aim of this article is to introduce the new Japanese context of concepts of people-centered early warning used
early-warning system to the international landslide community. In worldwide, the Japanese system couples monitoring and warning
this article, the method, the system, and the result of its application services with dissemination and communication.
to landslide disasters in 2009 are presented. In establishing an early-warning system with practical
application, it is important to: (1) establish the method for
Keywords Landslides . Early-warning system . Debris ows . estimating the locations and timing of landslides occurrences; (2)
Slope failures . Radial Basis Function Network determine the roles of the players; and (3) consider what kind of
information to disseminate using mass communication.
Introduction For establishing the method for estimating the location and
Countermeasures against landslides may be classied into timing of landslides occurrences using meteorological factors,
structural and nonstructural measures. Structural measures are many researchers have proposed a variety of methods, because
effective to prevent the occurrence of the hazard and to protect an many different factors inuence the occurrence of landslides.
area of livelihood from the hazard. Nonstructural measures are an Intense rainfall is the primary meteorological factor that is
important supplement to improve the effectiveness of structural recognized as directly associated with triggering landslides.
measures against landslide disasters. Antecedent precipitation is recognized as a secondary landslide-
In the Hyogo framework for Action, 20052015, which was triggering meteorological factor (for example, Wieczorek and
adopted by the World Conference on Disaster Reduction held Glade 2005). Wieczorek and Glade (2005) summarized the criteria
from 18 to 22 January 2005 at Kobe, Japan (UN ISDR 2005), early- for debris-ow triggering threshold used in various countries.
warning systems are recognized as important tools for disaster Their summary shows that various rainfall indices have been used
risk reduction and for achieving sustainable development and for determining the criteria. For example, Glade et al. (2000)
livelihoods. According to a United Nations global survey of early- shows that rainfall probability thresholds can be established by
warning systems (UN ISDR 2006), the object of people-centered applying an Antecedent Daily Rainfall model which uses two
early-warning systems is to enable individuals and communities variables: antecedent daily rainfall index and daily rainfall. A
threatened by hazards to act in sufcient time and in an method similar in concept to that of Glade et al. (2000) has been
appropriate manner so as to reduce the possibility of personal used in the Japanese early-warning systems that determine the
injury, loss of life and livelihood, and damage to property and the criterion of landslides occurrences using two rainfall indices.
environment. Such systems consist of four elements: risk knowl- In this article, the authors present an outline of the current
edge, monitoring and warning services, dissemination and landslide early-warning system which has been in operation since
communication, and response capability. 2007, operated by MLIT and the Japan Meteorological Agency
As part of the Japanese measures against landslide disasters, (JMA). Background of the new Japanese method of setting criteria
the Ministry of Land, Infrastructure, Transport and Tourism of early-warning information issue using rainfall indices presents a
(MLIT) has progressively installed structural measures such as brief history of Japanese early-warning systems so as to specify
catch dams and retaining walls in the valleys where the problems to be solved by the new method. The way to setting

Landslides 7 & (2010) 325


Technical Note

Fig. 1 Sketches of typical debris flow (left; after Varnes (1978) revised by the deep-seated landslide is a slow soil-block movement on a slip surface deeper than
authors), deep-seated landslide (middle; after Ohyagi (2007) revised by the 2 m below the ground surface, and a slope failure is a fast, shallow, soil
authors), slope failure (right; after Ohyagi (2007) revised by authors). In Japan, a movement on a slope steeper than 30
debris flow is a flow mixed with water, sand, and stones driven by gravity; a

criteria of early-warning information issue shows the theory and Based on these hypotheses, the underpinning issues are: (1)
technique of setting the criterion of disaster occurrence. Japanese to select appropriate rainfall indices; (2) to improve on the
early-warning system for debris ows and slope failures shows an method of discriminating the boundary between occurrence and
outline of the whole Japanese early-warning systems including how it non-occurrence rainfall; and (3) to collect locations and timing
operates, the ow of transmission of early-warning information, and of many rainfall-related mass-movement occurrences.
the format of early-warning dissemination. In Use of early-warning Figure 2 presents a sketch of the basic concept, showing that
information, some results of applying the method to actual disasters it is possible to draw various lines as the criterion of disaster
in 2009 are presented. Conclusions are presented in Conclusion and occurrence line (Critical Line, CL) depending on the method. A
future issues. linear CL is the easiest to set, but an arbitrary shaped CL seems
Note that in this article, the meaning of the term landslide differs to be the most precise discrimination line based on the data of
from its usual international meaning. The term landslide is widely occurrence and non-occurrence rainfall.
used internationally to represent the whole phenomena of mass Table 1 shows the history of various rainfall indices used by
movement; for example, the classication system as dened by MLIT (Terada and Nakaya 2001). These indices were revised
Varnes (1978). However, from the viewpoint of the Japanese according to analysis of records of mass-movement disasters
administration, these international classication systems cannot be and their associated rainfall. The methods before 2005 used a
used, because the Japanese law, dealing with prevention works that researcher or senior engineer to draw the Critical Line as a
include structural and nonstructural measures against disasters straight line tted by eye, because adequate records of disasters
related to mass movement, denes the types of landslides as (1) for statistical analysis do not exist in many regions.
debris ows, (2) slope failures, and (3) landslides in a limited sense
(deep-seated landslide). Also in the law, the types of material of these
three phenomena are dened as soil excluding rock (corresponding
to the international denition of engineering soil). Figure 1 shows the
typical diagrams of debris ows, slope failures, and landslides. Thus,
in this article, landslides means deep-seated landslide only, and mass
movement is used as the term representative of the broader
phenomenon.

Background of the new Japanese method of setting criteria


of early-warning information issue using rainfall indices

Basic concept
Since 1984, the basic concept for issuing early-warning informa-
tion in Japan has been based on two hypotheses which have not
changed. The rst hypothesis is that mass-movement occurrence
can be predicted using both a short-term rainfall index and a
long-term rainfall index, because mass movements are driven by Fig. 2 The basic concept used for setting the criterion for issuing early-warning
both surface water and ground water. The second hypothesis is information in Japan. The criterion of disaster occurrence is defined as the line
that, on the plane with short- and long-term rainfall index axes discriminating between the area of high and low probability of disaster
occurrence. This line has been termed the CL, which may be linear, curved, or
respectively, the area of mass-movement occurrence and non-
arbitrary shaped depending on the choice of the fitting method and is based on
occurrence can be identied by the plot of rainfall with disasters the distribution of occurrence rainfall (rainfall with disasters) and non-
(occurrence rainfall) and without disasters (non-occurrence occurrence rainfall (rainfall without disasters). This illustration omits many of
rainfall). the non-occurrence rainfalls from the low-probability area

326 Landslides 7 & (2010)


Table 1 Summary of rainfall indices and methods of setting the Critical Line used by MLIT in Japan
Year Short-term rainfall index Long-term rainfall index Method of boundary fitting
1984 60-min cumulative rainfall AP (half time: 24 h) By eye
1984 Effective rainfall AP (half time: 24 h) By eye
1993 AP (half time: 1.5 h) AP (half time: 72 h) By eye
2005 60-min cumulative rainfall Soilwater index Radial Basis Function Network
P
After Terada and Nakaya 2001, with the addition of the method of 2005 by the authors. AP is defined as P ai Ri , where P is AP, a 0:5i=T is decay coefficient before
ith hour, Ri is hourly reinfall before ith hour, and T is a relaxation time (time to halve) (h)
AP antecedent precipitation

Problems with previous methods and merits of the new method investigating the merit of implementing this method in the early-
The method since 2005 follows the above hypotheses but solves the warning information system for mass-movement disasters. In
following problems which were revealed in the previous methods: 2005, the new early-warning system was established.

1. The previous methods needed a vast amount of mass-move- The way to setting criteria of early-warning information issue
ment-occurrence rainfall. The main contents of this section are based on the manual
2. No records of mass-movement occurrences exist for many describing at the methodology to set the criterion written by the
regions. Department of Erosion and Sediment Control, MLIT, JMA, and
3. A subjective tting technique to set the criterion generates an NILIM (DESC 2005) and on the study of Kuramoto et al. (2001,
accuracy bias from one region to another. 2005) and Okada (2005). The objective of the new early-warning
4. The spatial density of rainfall gauges is insufcient to predict CLs system in Japan is to warn of debris ows and slope failures with
precisely for every region in Japan, especially in the mountains. high spatial frequency. The main procedure for setting the criterion
for the occurrence of debris ows and slope failures in each 5-km
The range of potentially useful rainfall information has been grid mesh covering on Japan is shown in Fig. 3. The currently
developing. JMA has produced 2.5-km grid mesh rainfall data since adopted short-term rainfall index is the 60-min cumulative rainfall,
2001 (5-km grid mesh from 1988) known as Radar Automated and the long-term rainfall index is the soilwater index. The
Meteorological Data Acquisition System analytical rainfall (Radar investigation of the choice of rainfall indices was researched by
AMeDAS analytical rainfall) and short-time forecasts of rainfall (from Kuramoto et al. (2005) and is discussed in Rainfall indices.
1 to 6 h; hereafter actual rainfall and forecast rainfall, respectively).
These actual rainfall data are provided by estimating rainfall
intensity with radar checked against gauged AMeDAS data. Forecast The theory of Radial Basis Function Network
rainfall is more accurate in the nearest future. That is, forecast rainfall As mentioned in the previous section, the new method was
in the next hour is more reliable than that forecast in 6 h. These developed to overcome problems such as the subjective tting
rainfall data are available to solve the 4th problem mentioned above.
Kuramoto et al. (2001) researched a method to solve the 1st to 3rd
problems mentioned above. The method is based on the following
fundamental concepts:

1. The target mass movements are debris ows, and slope


failures with high spatial density, except landslides.
2. The criterion is calculated based on two rainfall indices: a
short-term rainfall index and a long-term rainfall index.
3. An arbitrary shaped CL can be drawn objectively using only
non-occurrence rainfall with Radial Basis Function Network
(RBFN).
4. The shape of CL is easily revised as new data come to hand.

Using the method of Kuramoto et al. (2001), the concept of


setting CL can be changed from perceiving CL as the boundary
between areas of low and high probability of disaster occurrences
that depend on occurrence and non-occurrence rainfall plot
(Fig. 2), to specifying the area of low probability using only the
non-occurrence rainfall.
Fig. 3 Procedure for setting the critical-line criterion for the occurrence of debris
This idea is attractive, and the result is more precise and flows and slope failures. A Critical Line is determined for each 5-km grid mesh
objective than MLITs previously used methods. MLIT started block

Landslides 7 & (2010) 327


Technical Note
of the criterion and the paucity of existing mass-movement shown in Eq. 2, represented as the total of the products of the load
records in some regions. In this subsection, the RBFN theory is coefcients and the output of the hidden-layer elements.
described. X
m
RBFN is an approach to designing a supervised neural network as Ox wj hj x 2
an approximation problem in polydimensional space. This theory is j1
applied to problems such as recognition and interpolation (for
where, O(x) is the output value of output layer element, and wj is
example, Er et al. 2002). The learning of the network is equivalent to
the load coefcient. RBFN output value described hereafter is O(x).
nding a surface in multidimensional space that provides a best t to
It is important to accurately reproduce the dataset. With
the observed (training) data through statistical criteria. Correspond-
p sets of training data xi ( i=1, 2,, p) and desired output data
ingly, generalization is equivalent to using this multidimensional
(supervised data) yi (i=1, 2,, p; here, we set yi =1) and m sets of
surface to interpolate the test data (for example, Orr 1996; Haykin
the basic functions, the learning of RBFN is performed to calculate
1999). This network consists of three layers that are the input layer (n
the optimum w=wj ( j=1, 2, . . . , m) to minimize the sum of the
elements), the hidden layer (m elements), and output layer (one
squared difference between the output values using training data O
element) as shown in Fig. 4. Each layer consists of elements that
(xi) and the desired data shown in Eq. 3.
perform input and output. Elements in each layer are linked between
layers for transmitting the information (the elements within each layer X
p X
m

are not linked). The transmission of information in RBFN is one-way E yi  Oxi 2 j w2j ! Min 3
i1 j1
from the input layer to the output layer. The resulting surface of the
RBFN interpolation of the test data is called response surface. where j is the regularization parameter (detailed in Determi-
In the process of transmitting information from input layers to nation of response surface using RBFN).
the hidden layer, a Gaussian function in Eq. 1 is used as the radial To calculate optimum w, the partial derivative of E in the Eq. 3
basis function for the output of the jth element in the hidden layer with respect to wj ( j=1, 2,, m) is calculated as:
using data from input elements linked to that element.
  ! @E Xp
@Oxi
x  cj 2
hj x exp  1 2 Oxi  yi 2j wj 4
r2 @wj i1
@wj
@Oxi
Thus, according to @wj hj xi , Eq. 4 is summarized as:
where, hj (x) is output data from jth element in the hidden layer,
x(=x1,,xn) is input data vector from input elements linked to jth X
p X
p

element in the hidden layer, cj is the center point of the basis function Oxi hj xi j wj yi hj xi : 5
i1 i1
and r is radius of the basis function. 60-min cumulative rainfall and
soilwater index are used as input data vector in our case. If cj =cj1, Eq. 5 can be rewritten in the matrix form as:
cj2,, cjn is assumed, it is possible P for the 
 basis function to be
2
an ellipse of the form: hj x exp  nk1 xk  c jk =rk2 . r can be HT O LwHT y 6
determined arbitrary or experimentally.
Where
A weighting factor forms part of the connecting line between
elements of the hidden layer and the output layer and is called a load 2 3 2 3 32
coefcient. So, the output value of the output layer element is, as Ox1 1 0 ... 0 1
6 Ox2 7 60 2 ... 0 7 6 2 7
6 7 6 7 6 7
O 6 .. 7; L 6 .. .. .. ..7; y 6 .. 7
4 . 5 4 . . . .5 4 . 5
 
O xp 0 0 ... m p

and H is called the output matrix in the hidden layer,


2 3
h1 x1 h2 x1  hm x1
6 h1 x2 h2 x2  hm x2 7
6 7
Hh1 h2    hm  6 .. .. .. .. 7:
4 . .  5
  .  .
h1 xp h2 xp  hm xp

Here, the output value of the output layer can be describe as


O Hw. Substituting this expression into the Eq. 6 gives:
HT y HT OLw
HT HwLw
 
HT HL w

then the optimum w can be solved as:


 1
w HT H L HT y: 7
Fig. 4 The three-layer design of the RBFN

328 Landslides 7 & (2010)


Fig. 5 Output of RBFN using a test
dataset of rainfall indices. Left Three-
dimensional view of the output
response surface. The maximum value
is 1.0 because all of the 0 mm rainfalls
with 0 mm of soil water are non-
occurrence rainfalls. Right Contoured
two-dimensional plot (contour lines at
0.1 intervals a potential candidates for
the critical line) of response surface on
60-min cumulative rainfall and soil-
water index as x and y axes,
respectively

The learning of RBFN is equivalent to solving the inverse matrix surface reects the characteristics of rainfall of the region.
A1 =(HT H+)1. This feature of RBFN learning is effective in The response surface represents a probability density func-
adding data for additional learning because learning can be tion of non-occurrence rainfall. The value of the response surface
completed by updating the inverse matrix A1. That also means is the RBFN output value as z-axis (Fig. 5). On the surface, a high
that learning is completed using only input data such as non- RBFN output value means a low probability of disaster occur-
occurrence rainfall. rences because controls the output value in inverse proportion
to the amount of data in the lattice cell.
Determination of response surface using RBFN
The output value of RBFN uses the non-occurrence rainfall to Rainfall indices
create the response surface of the grid. Non-occurrence rainfall data The development of indices of rainfall was an essential part of
has been collected for each 5-km grid mesh at 1-h time resolution for establishing the Japanese early-warning systems.
more than 10 years created by JMA. 60-minute cumulative rainfall
and soilwater index are calculated for each hour using the collected Soilwater index
data. In the following, the method for determining the response The Soilwater index (Okada 2005) is a concept model that uses a
surface in each grid mesh using RBFN is described (Fig. 5). calculated value of the total water depth of a three-layer tank
Usually in RBFN, the response surface is calculated by super- model estimated using xed parameters for each 5-km mesh grid
position of basic functions which are set for each data point. But mesh in Japan (Fig. 7).
where there is a very large amount of data, determining basis Tank model is a hydrological model usually used to
functions using all of the data to make learning complete, leads to a calculate water discharge. Its absolute value is important in
calculated response surface which may have a low generalization
capacity because of the unbalanced distribution of data. It is
important to reduce the number of basic functions and to equalize
the frequency density of data so as to create a response surface with
high generalization capacity.
To reduce the number of basic functions, the basic functions are
set for the corners of a lattice at constant intervals (Rx, Ry),
summarizing all of the rainfall data that exists inside the lattice cells as
shown in Fig. 6. To equalize the density of data over the surface, the
rainfall data are clustered using the centroid method for each of the
hachured lattice cell shown in Fig. 6. This single clustered piece of data
for each lattice cell is the data used in the input layer. The reliability of
each basis function varies depending on the amount of data. So the
regularization parameter in Eq. 3 is set to reducing this distortion. is
given as:

max  min
min 8
1 count
where max is the maximum value of , min is the minimum
value of , count is the amount of data in lattice cell, for example
the data in cluster 1 in Fig. 6. The value of max and min can be
determined arbitrary or experimentally under the condition
max min.
controls the output value O(x) which tends to be small if
there are few data in the lattice cell. That means that the response Fig. 6 The method of clustering the rainfall data

Landslides 7 & (2010) 329


Technical Note
the mass movement in relationship to rainfall intensity and
antecedent rainfall in the region. According to the research of
Okada (2005), the rank order of the soilwater index remains
unchanged in many cases when the parameters of the tank
model are changed. The xed parameters identied by statistical
analysis of the relationship between rainfall and discharge in a
granite region in Japan listed in Table 2.
In the new Japanese early-warning information method, the
absolute value of the soilwater index is used as x-axis. The
RBFN output value, however, is computed relativity which has
basically the same meaning as the rank order of the soilwater
index, because the RBFN output value is calculated from much
past non-occurrence rainfall data which represent the character-
istics of the rainfall in the region. in Eq. 8 makes extreme
value smooth. This feature of RBFN is the solution of the
problem of rank order of soilwater index, that if extreme value
is observed, the value is registered as high rank order of the
region. Thus RBFN is also methodology that includes the merit
of soilwater index and the solution of the problem of soil
water index.

Fig. 7 Diagrammatic representation of the three-layer tank model. The absolute Selection of combination of rainfall indices
value of the soilwater index is defined as the sum of water depths in each of the
The investigation of appropriate short- and long-term rainfall
layers. The parameters determining the soilwater index are shown in Table 2
indices for Japanese new early-warning system was conducted by
Kuramoto et al. (2005). To investigate various combinations of
rainfall indices shown in Table 3, response surfaces were created
most usage. However the absolute value of the soilwater index using rainfall data from 1994 to 2000 from which all of the
is not so important in predicting the occurrence of mass occurrence rainfalls had been excluded. The CL was then selected
movements according to Okada (2005). The more important using the records of 1,350 debris ows and slope failures known to
factor is the relative rank order of the current value of soil have occurred in that interval (Determination and revision of the
water index in the region under torrential rainfall conditions, critical line presents details of how the CL is determined).
comparing to the previous largest value and 2nd largest value Three numbers were calculated to assess the reliability of the
and so on during 10 years in the same region. The rank order of selected rainfall indices in discriminating the selected hazard
the soilwater index includes the concept of recursive time of occurrences: a disastercapture ratio, the frequency of false

Table 2 Parameters of the soilwater index


Tank Fist Second Third
Outflow height (mm) L1 =15 , L2 =60 L3 =15 L4 =15
Outflow coefficient (1/h) 1 =0.1, 2 =0.15 3 =0.05 4 =0.01
Coefficient of permeability (1/h) 1 =0.12 2 =0.05 3 =0.01

Table 3 The various combination of rainfall indices investigated by Kuramoto et al. (2005)
Case Short-term rainfall index Long-term rainfall index
Case 1 10-min cumulative rainfall Soilwater index
Case 2 30-min cumulative rainfall Soilwater index
Case 3 60-min cumulative rainfall Soilwater index
Case 4 120-min cumulative rainfall Soilwater index
Case 5 180-min cumulative rainfall Soilwater index
Case 6 60-min cumulative rainfall AP (half time: 24 h)
Case 7 AP (half time: 1.5 h) AP (half time: 72 h)
AP antecedent precipitation

330 Landslides 7 & (2010)


Table 3. Using these analyses, Kuramoto et al. (2005) concluded
that Case 3 was the best combination for an efcient early-warning
system using RBFN, because the disastercapture ratio was
relatively high, and the false frequency was low (Fig. 8), and long
mean exceed times were avoided (Fig. 9).

Determination and revision of the critical line


To provide an objective criterion to be met before evacuation
instructions and evacuation orders are disseminated, it is
necessary to choose a particular isovalue line on the response
surface as the CL for the region. Once a CL has been calculated by
RBFN, its shape must be revised to avoid any contradictions from
the actual physical phenomenon.
The following subsection explains the method by which the
Fig. 8 The relationship between the disastercapture ratio and the false
frequency (Kuramoto et al. 2005) CL is revised. After these processes, the nal CL is available for
governmental operation in the region.

Procedure for CL revision


Because the soilwater index is the value of the sum of depths
from each tank layer, the value of the 60-min cumulative rainfall
is always added to the rst tank at the time. Hence, the portion of
the graphical space where the soilwater index is less than or
equal to the 60-min cumulative rainfall is unreal and must be
ignored in the analysis by treating it as unreal.
In addition, the gradient of the real CL can only take a
limited range of values (0) such that the slope angle of the
CL lies between 0 and 90 in graphical space, and any
computed value outside of this range is unreal and must be
eliminated from the functional CL before it is implemented for
the region. To achieve this in cases where the 60-min cumulative
Fig. 9 Relationship between disastercapture ratio and mean exceed time rainfall at points such as points A2 and A3 are less than the value
(Kuramoto et al. 2005) of point A1 (Fig. 10), the critical soilwater index of A2 and A3 are
adjusted so they are the same value as at point A1.
recognition and a mean time to exceed the CL (termed exceed In cases where soilwater index is less at points B2 and B3
time). The disastercapture ratio is dened as ratio between than the value at point B1 (Fig. 11), 60-min cumulative rainfall of
number of occurrence rainfalls occurring above the CL threshold B2 and B3 are adjusted so they are the same as the 60-min
using selected CL and the total number of occurrence rainfalls. cumulative rainfall at point B1.
The false frequency is dened as the frequency of occurrence In cases where the gradient of CL becomes negative (Fig. 12),
rainfalls below the selected CL threshold. The mean exceed time is soilwater index of C2 is adjusted so that it is equal to soilwater
dened as the mean time for the trace of the rainfall indices to drift index of point C1.
over the selected CL threshold. Figure 8 shows the relationship These procedures are necessary because, although the
between the disastercapture ratio and the false frequency, and calculated best-tting surface may suggest that there is more
Fig. 9 shows the relationship between the disastercapture ratio and than one critical 60-min rainfall intensity threshold associated
mean exceed time for the variety of tested indices listed in with a given soilwater index, this is unreal and is an artifact

Fig. 10 The method of revising the


critical line with respect to the soil
water index. The figures show before
(left) and after (right) revision

Landslides 7 & (2010) 331


Technical Note
Fig. 11 The method of revising of the
critical line with respect to the 60-min
cumulative rainfall. The figures show
before (left) and after (right) revision

Fig. 12 The method of revising the


gradient of the critical line. The
figures show before (left) and after
(right) revision

Fig. 13 The method for setting a


lowest limit of soilwater index. The
figures show before (left) and after
(right) revision

332 Landslides 7 & (2010)


Fig. 14 Left the critical line is
revised based on a past large
disasters. Right the critical line is
revised based on operational
experience to take account of the
value of historical occurrence rainfall,
evacuation instruction, or order

Fig. 15 Critical line candidates on an


already-revised response surface (left).
The trace (snake line) of a
particularly large rainfall event that
triggered slope failure (occurrence
rainfall; right)

Fig. 16 Relationship between RBFN output value of isovalue (contour) line and
disastercapture ratio. Legend indicates dates of large rainfall events with
number of slope failures (in parenthesis); for example, 50 disasters occurred on
16th of September 1991 Fig. 17 The methodology of setting the lowest limit of the soilwater index

Landslides 7 & (2010) 333


Technical Note
Fig. 18 Basic concept of the currently
operating Japanese early-warning
system discussed in the paper

of the tting procedure and the data that the surface have been rience (Fig. 14). Such revision is indicated because the CL being
tted to. used does not reect the true characteristics of rainfall
False predictions will frequently occur in the region, when condition of the region.
rainfalls of high intensity for short periods, such as in a squall,
occurs after a period without rainfall (Fig. 13). These false Technique of CL selection and determining the lowest limit of the soilwater
predictions can be prevented by setting a lowest limit to the index
soilwater index. The way of setting this limit is described in If non-occurrence rainfall data can be collected, the response
Technique of CL selection and determining the lowest limit of surface can be created. However, to select CL and to determine
the soilwater index. the lowest limit of soilwater index, it is better to use records of
Once a CL for implementation has been initially deter- debris-ow and slope-failure occurrences. If there are no
mined, it may be necessary to revise the CL after the recorded disasters in the particular region, its CL can be
occurrence of large disasters and following operational expe- determined as 0.1 because CL should include almost all of the

Fig. 19 Basic concept of the monitoring


sequence leading to dissemination of
early-warning information

334 Landslides 7 & (2010)


Fig. 20 The role of major players in transmission of early-warning information

non-occurrence rainfalls so that the area of low probability of Japanese early-warning system for debris flows and slope failures
disasters is maximized. A nationwide initiative for dissemination of early-warning
Various candidates for selection as the CL (0.1 interval) on a information was implemented at the end of March 2008, after
response surface are shown in Fig. 15 (left), and next to it is the the method was initiated in 2005. This system is aimed at
selection of the trace of the history of the accumulating rainfall and facilitating the evacuation of residents in advance of the
changing soilwater index (here referred to as the snake line; right; occurrence of disasters, and at assisting the decision-makers such
see Japanese early-warning system for debris ows and slope as mayors to judge the timing of when to disseminate evacuation
failures for detail) during a rainstorm that lead to a particular large instructions or orders. Figures 18 and 19 summarize the whole
occurrence rainfall event. The candidates for CL have already been system.
revised by the procedures shown in Figs. 10, 11, and 12. The main players who send out early-warning information
The relationship between RBFN output values of the to the residential population are JMA and local government
various candidates for CL and their disastercapture ratios is (Fig. 20). When torrential rain is expected or falling, the timing
shown graphically in Fig. 16. This is the methodology used for of the issuing of early-warning information is determined by
selecting CL. If the CL is selected as the line which has an the expected values of the 60-min cumulative rainfall and soil
RBFN output value of 0.4, almost all of the collected past water index calculated using the forecast rainfall for 13 h into
disasters in the region will be recognized as events calling for the future. The progress of the actual values of the two indices
issuing a warning. is logged graphically as a snake line in the graphical space of
The methodology for determining the lowest limit of soil Fig. 21 so that the likelihood of exceeding the CL in the near
water index is illustrated in Fig. 17. First of all, a target disaster future can be anticipated to provide enough lead time to
is selected. These are dened as disasters caused by a relatively evacuate residents before the actual rainfall causes the CL to be
large historical occurrence rainfall in that region. All of the
grid blocks in that region are then searched for the minimum
soilwater index that exceeds the selected CL at the location of
the disaster. In the grid block, the ratio of the lowest limit of
soilwater index is calculated using x/xmax (as indicated in
Fig. 17). Finally, the value in each grid block which is calculated
as the maximum soilwater index on the CL for the grid block
is multiplied by this ratio to dene the lowest limit of the soil
water index for the grid block.

Fig. 21 The evolution of the snake line under actual operational conditions,
with projection using forecast rainfall over the next 13 h Fig. 22 The format in which early-warning information is disseminated from JMA

Landslides 7 & (2010) 335


Technical Note
exceeded. This allows JMA to initiate the early-warning of as in Fig. 23). The result of this information, of course, is the
debris ows and slope failures. The weather news on a TV, same as JMAs result, but the degree of danger of each grid
radio, and the Internet then deliver the early-warning infor- block can be shown. The aim of this expression is to provide
mation. ner information for the local residents and decision-makers
An example of the format in which the information is of the region.
delivered by JMA is shown in Fig. 22. This report says which
local government precincts are now considered to be in a Use of early-warning information
dangerous situation and which are now released from the To illustrate the use of the early-warning system, the results of its
warning. This precinct-level information is calculated using the operation in 2009 are provided. The relationship among rainfall
mean value or maximum value of 5-km grid meshes inside the conditions, the time of occurrence of debris ows and slope
boundary of the governmental precincts. Local governments failures, and the times of dissemination of the early-warning
send the early-warning information in a different format (such information are shown.

Fig. 23 The format in which early-


warning information is disseminated
from local government (top). The
colors are coded to show the
condition indicated by the danger
level, and the action that is
recommended be taken (below)

336 Landslides 7 & (2010)


Fig. 24 Left The time series of 60-min cumulative rainfalls and soilwater index information was in effect. Right Snake line based on the rainfall is depicted
for the occurrence event of 21st July 2009, together with the times of occurrence with timing of disasters occurrences. This figure shows that the critical line is
of debris flows and slope failures, and the duration for which the early-warning adequate and that the early-warning information is effective

On 21st July, 2009, 65 debris ows and 105 slope failures Kyoji Sassa for encouraging us to write this article and giving us
occurred in Yamaguchi Prefecture, and MLIT report that 14 helpful advice.
people died as a result of these mass movements.
The time series of 60-min rainfalls and soilwater index, the
times of occurrence of debris ows and slope failures, and the References
period of early-warning information issue are shown in Fig. 24
(left). The progress of the snake line and the timing of the Department of Erosion and Sediment Control (DESC), Ministry of Land, Infrastructure,
Transport and Tourism (MLIT) and Japan Meteorological Agency (JMA) and National
disasters (red box) in this grid block are shown in Fig. 24 (right). Institute for Land and Infrastructure Management (NILIM), (2005). Manual for the
Figure 24 indicates that CL adequately captured the timing of method of setting mass-movement disasters warning criterion based on rainfall
these disasters occurrences. indices (Draft), p.21 (in Japanese)
Er MJ, Wu S, Lu J, Toh HL (2002) Face Recognition With Radial Basis Function (RBF)
Conclusion and future issues Neural Networks. IEEE Trans Neural Netw 13(3):697710
Glade T, Crozier MJ, Smith P (2000) Applying probability determination to refine
A new early-warning system was established in Japan in 2005
landslide-triggering rainfall thereshols using an empirical Antecedent Daily Rainfall
and has operated since 2007. The main methodology is to set a Model. Pure Appl Geophys 157(6/8):10591079
Critical Line (danger threshold) with RBFN using the rainfall Haykin S (1999) Neural networks: a comprehensive foundation 2nd ed. Prentice-Hall,
indices of 60-min cumulative rainfall and calculated soilwater Inc., pp 256-257
index. The merit of this method is the ability to apply it to Kuramoto K, Noro T, Osanai N, Kobayashi M, Okada K (2005) A study on rainfall indexes
for giving early warning information for sediment-related disasters, Proceedings of
areas with no prior record of disasters. Updating is easily
Annual Research Meeting in 2005, pp.186-187, Japan Society of Erosion Control
accomplished because the system learns from the input data. Engineering (in Japanese)
The latter characteristic means that the criteria can be revised Kuramoto K, Tetsuga H, Higashi N, Arakawa M, Nakayama H, Furukawa K (2001) A
easily while the system is operating. study on a method for determining non-linear critical line of slope failures during
It is difcult to conclude that the available short-term heavy rainfall based on RBF network, Doboku Gakkai Ronbunshu, No.672, VI-50,
pp.117-132, Japan Society of Civil Engineers (in Japanese with English Abstract)
forecast rainfall that is used to anticipate the level of danger is
Ohyagi N (2007) Air photograph interpretation for analysis on landslide structures, p
accurate enough in time and space for this purpose. This is 320, kinmiraisha (in Japanese)
because mass-movement disasters represent phenomena trig- Okada K (2005) Soil water index. Sokkoujihou No.69, 5(5):67-100 (in Japanese)
gered by complex interactions of various factors, and the use Orr MJL (1996) Introduction to radial basis function networks. Technical report,
of rainfall indices alone for predicting their occurrences has Institute for Adaptive and Neural Computation, Division of Informatics,
Edinburgh University
serious limitations.
Terada H, Nakaya H (2001) Operating methods of critical rainfall for warning and
However, the result from applying the system to debris- evacuation from sediment related disasters, Technical note of National Institute for
ow and slope-failure disasters in 2009 suggests that the Land and Infrastructure Management No.5, National Institute for Land and
Japanese new early-warning system has powerful potential to Infrastructure Management, p 71 (in Japanese with English Abstract)
reduce injuries from debris ows and slope failures if the CL is United Nations Inter-Agency Secretariat of the International Strategy for Disaster
Reduction (UN ISDR) (2005) Hyogo Framework for Action 2005-2015: Building the
set adequately using many non-occurrence rainfalls and
resilience of nations and communities to disasters, Extract from the final report of
records of disasters. the Wold Conference on Disaster Reduction (A/CONF.206/6) p 22
United Nations Inter-Agency Secretariat of the International Strategy for Disaster
Acknowledgement Reduction (UN ISDR) (2006) Global survey of early warning systems: an assessment
We would like to thank gratefully Dr. Mauri McSaveney for giving of capacities, gaps and opportunities toward building a comprehensive global early
warning system for all natural hazards, p 46
us constructive advice to make this article readable and clear. Varnes DJ (1978) Slope movement types and processes. In: Schuster RL, Krizek RJ (eds)
Anonymous reviewers give us constructive comments to improve Landslides: analysis and control, National Research Council, Transportation Research
the contents of this article. Finally, we would like to thank Prof. Board, Special Report 176, pp. 11-33

Landslides 7 & (2010) 337


Technical Note
Wieczorek GF, Glade T (2005) Climatic factors influencing occurrence of debris flows. In: 1-6, Minamihara, Tsukuba, Ibaraki Prefecture, Japan
Jakob M, Hunger O (eds) Debris-flow Hazards and Related Phonomena, Praxis, e-mail: t-shimizu@pwri.go.jp
Springer Berlin Heidelberg, pp. 325-362
K. Kuramoto
Chuden Engineering Consultants Co., Ltd,
2-3-30 Deshio, Minami-ku, Hiroshima, Hiroshima Prefecture, Japan
e-mail: kkuramot@cecnet.co.jp
N. Osanai
Erosion and Sediment Control Division, S. Kojima
National Institute for Land and Infrastructure Management (NILIM), MLIT, Fujikawa Sabo Office,
1, Asahi, Tsukuba, Ibaraki Prefecture 305-0804, Japan MLIT,
e-mail: osanai-n92ta@nilim.go.jp Fujimi, Japan

T. Shimizu ()) T. Noro


Volcano and Debris Flow Research Team, Snow Avalanche and Landslide Research Center, PWRI,
Public Works Research Institute (PWRI), 2-6-8 Nishiki Town, Myoukou, Niigata Prefecture, Japan

338 Landslides 7 & (2010)