Transportation Research Part E 43 (2007) 737754

www.elsevier.com/locate/tre

A scenario planning approach for the ood emergency

logistics preparation problem under uncertainty
a,*
Mei-Shiang Chang , Ya-Ling Tseng b, Jing-Wen Chen c

a
Department of Civil Engineering, Chung Yuan Christian University, Chung-Li 32023, Taiwan
b
Department of Civil Engineering, National Central University, Chung-Li 32001, Taiwan
c
Institute of Business Administration, Chung Hua University, Hsin-Chu 30067, Taiwan

Received 25 January 2006; received in revised form 21 September 2006; accepted 5 October 2006

Abstract

This paper aims to develop a decision-making tool that can be used by government agencies in planning for ood emer-
gency logistics. In this article, the ood emergency logistics problem with uncertainty is formulated as two stochastic pro-
gramming models that allow for the determination of a rescue resource distribution system for urban ood disasters. The
decision variables include the structure of rescue organizations, locations of rescue resource storehouses, allocations of res-
cue resources under capacity restrictions, and distributions of rescue resources. By applying the data processing and net-
work analysis functions of the geographic information system, ooding potential maps can estimate the possible locations
of rescue demand points and the required amount of rescue equipment. The proposed models are solved using a sample
average approximation scheme. Finally, a real example of planning for ood emergency logistics is presented to highlight
the signicance of the proposed model as well as the ecacy of the proposed solution strategy.
2007 Elsevier Ltd. All rights reserved.

Keywords: Emergency facility location; Rescue resource allocation; Scenario planning; Geographic information system

1. Introduction

Any occurrence of natural disasters, such as earthquakes, typhoons, oods, or drought, will cause huge
property damage and human injuries. In many countries, the ood disaster has the highest occurrence fre-
quency among all natural disasters. Compared with earthquakes, ood disasters are more easily predicted
and prevented. Therefore, it is very important for ood disaster management teams to make a prudent and
comprehensive plan for ood disaster prevention. Consequently, government agencies are spending tremen-
dous amounts of money and manpower for ood disaster prevention and rescue.
In urban areas, drainage often overows due to the large quantity of water that pours into drains over a
short time period. Consequently, the underground drainage pipes may get blocked resulting in ooded areas

*
Corresponding author. Tel.: +886 3 2654210; fax: +886 3 2654299.
E-mail address: [email protected] (M.-S. Chang).

1366-5545/$ - see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.tre.2006.10.013
738 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754

that are relatively dicult to predict precisely. Thus, rescue demand is full of uncertainties. In addition, once
rescue demand is requested, rescue equipment and relief personnel must be transported eciently to minimize
the damage in aected areas. At the preparation planning stage of ood emergency logistics, the authorities
must consider all possible circumstances in advance and make prudent and sedulous emergency logistics plans
to distribute the rescue equipment.
Currently, such comprehensive ood preparation planning tools currently do not exist. This paper aims
at developing a decision-making tool which can potentially be used by disaster prevention and rescue agen-
cies for planning ood emergency logistics preparation. For this purpose, this study will take the practi-
cal issues of emergency logistics and the current organizational structure of disaster prevention agencies
into consideration. In addition, the disaster rescue demand point and the amount of demand are stochas-
tic. The suitable rescue resource preparation plan must be determined prudently by considering all
possible rescue demand scenarios. In this paper, we will explore how to optimize the expected perfor-
mance over all scenarios and incorporate scenario planning into ood emergency logistics preparation
modeling.
The paper is organized as follows. Section 2 reviews studies related to emergency logistics planning and
multi-echelon location models under uncertainty. In Section 3, two stochastic programming models are pro-
posed that oer suitable structures of rescue organization, storehouse locations, rescue resource allocations,
and distribution plans. In Section 4, we develop a heuristic to solve the problem presented in Section 3. This
is followed by computational results of model testing in Section 5. Conclusions and directions for future
research are provided at the end of the paper.

2. Literature review

2.1. Emergency logistics planning

Although logistics management literature is abundant, only some research addresses emergency logistics.
O zdamar et al. (2004) point out that emergency logistics planning involves dispatching commodities (medical
materials, specialized rescue equipment, rescue teams, etc.) to distribution centers in aected areas as soon as
possible to accelerate relief operations.
Ray (1987) develops a single-commodity, multi-modal network ow model on a capacitated network over a
multi-period planning horizon. The sum of incurred costs is minimized for the transportation and storage of
food. Knott (1987) proposes a linear programming model for the bulk food transportation problem in which
ecient use of the truck eet minimizes transportation cost and maximizes the amount of food delivered.
Knott (1988) adopts a linear programming model to determine vehicle schedules for transporting bulk food
to a disaster area.
In the context of disaster relief operations, Haghani and Oh (1996) present a multi-commodity, multi-
modal network ow problem with time windows as a single objective linear programming model on a
time-space network. Detailed routing and scheduling of available transportation modes, delivery schedules
of various commodities, and load plans for each transportation modes are determined. The objective is to min-
imize the sum of vehicle ow costs, commodity ow costs, supply or demand carryover costs, and transfer
costs over all time periods.
Fiedich et al. (2000) propose a dynamic combinatorial optimization model and a heuristic to determine an
optimized resource schedule for assigning resources in space and time to the aected areas after strong earth-
quakes, where the objective function is to minimize the total number of facilities during the search-and-rescue
period.
O zdamar et al. (2004) point out that the emergency logistics plan can be obtained by repetitively solving the
dynamic time-dependent transportation problem. The objective aims to minimize unsatised demands of all
commodities throughout the planning horizon. The emergency logistics plan involves optimal pick-up and
delivery schedules for vehicles within the considered planning time horizon as well as optimal quantities
and types of loads picked up and delivered on these routes. Those plans are regenerated while the new demand
requests occur or while new suppliers and transportation modes become available during the planning time
horizon.
 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754 739

It is clear that most of the literature on emergency logistics emphasizes how to dynamically distribute lim-
ited rescue resources to an area after a disaster has occurred. Planning for ood emergency logistics prepara-
tion has received little attention. To our knowledge, none of the studies used stochastic programming to deal
with this issue. As mentioned above, we will apply stochastic programming theory to determine ood emer-
gency logistics preparation.

2.2. Stochastic facility location problem and distribution network design

A large number of deterministic optimization-based approaches have been proposed for the design of sup-
ply chain networks (Aikens, 1985; Georion and Powers, 1995; Vidal and Goetschalckx, 1997). There is also a
great deal of research in the facility location component of supply chain network design under uncertainty
(Owen and Daskin, 1998). However, few studies have addressed the comprehensive design of supply chain net-
works under uncertainty. Researches on more complex multi-echelon location models under uncertainty only
appear in the literature of the past decade.
Gutierrez et al. (1996) propose a robust optimization framework for network design under uncertainty.
This approach seeks network congurations that are nearly optimal for a variety of scenarios of design param-
eters. A modied Benders decomposition algorithm is adopted to generate robust designs. A qualitative dis-
cussion of global supply chain design has been given by Vidal and Goetschalckx (1997, 2001), who present a
large-scale MIP for choosing plant locations and suppliers that incorporates the suppliers reliability into the
constraints.
MirHassani et al. (2000) formulate the supply chain network design problem as a stochastic program with
xed recourse. The rst-stage decisions, which involved opening and closing plants and distribution centers
and setting their capacity levels, must be decided prior to the realization of future demands. The objective
function coecients are deterministic; uncertainty is presented only in the right-hand sides of the recourse con-
straints, which may represent demands or capacities. The Benders decomposition algorithm is used to solve
the resulting stochastic integer program.
Tsiakis et al. (2001) consider a two-stage stochastic programming model for a multi-product, multi-echelon
supply chain under scenario-based demand uncertainty. The objective is to determine the middle-echelon facil-
ity locations and capacities, transportation links, and distribution ows to minimize the expected cost. The
model was formulated as a large-scale, mixed-integer linear programming model and was solved using
CPLEX.
Santoso et al. (2005) propose a stochastic programming model to formulate a global supply chain network
design problem with random costs, demands, and capacities. The problem is to decide where to build facilities
and what machines to build at each facility in order to minimize the total expected cost. A shortage penalty is
included in the total expected cost in case the constructed capacity is insucient to meet the realized demand.
The sample average approximation was integrated with an accelerated Benders decomposition algorithm to
quickly compute high quality solutions to large-scale stochastic supply chain design problems with many pos-
sible scenarios.
Although some published supply chain network design models and the described emergency logistics prep-
aration planning problem have some common characteristics, they cannot be adopted to specic ood condi-
tions for two reasons: the rescue demand points in a ood disaster occur randomly, and the distribution of
rescue resources requires mutual assistance among multi-group, multi-echelon, and multi-level suppliers. Nei-
ther of these factors is considered in the current literature; therefore, a new and more adequate stochastic
multi-backup location model must be developed.

3. Model development

Due to the uncertainty of ood disasters and the quick response requirement, authorities must create a
ood disaster response plan or standard operation procedure and reserve enough rescue equipment, resources,
and rescue teams at proper places in advance. Once the disaster has occurred, the rescue operation will be
launched into schedule without chaos. Basic ood rescue operations follow guidelines that divide the disaster
area into several groups and emphasize intra-group distribution and inter-group backup, despite the various
740 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754

structures of rescue organizations among dierent counties. Under such a disaster rescue system, the ood
emergency logistics network has a multi-group, multi-echelon, and multi-level structure. The structure of res-
cue organizations and the relation of mutual backup are illustrated in Fig. 1. As shown in the gure, nodes are
classied into ve types: a head rescue center, regional rescue centers, local rescue bases, contractors, and
demand points. Regional rescue centers, local rescue bases, and demand points can be further classied
according to the district in which they are located. In addition, there are four types of backup activities, which
are shown as global links, intra-group inter-level links, intra-group intra-level links, and inter-group links in
Fig. 1. The structure of rescue organizations and the ood rescue division will now be explained in detail.
According to the size of territory, rescue centers or storehouses are classied into several echelons, such as
head rescue center, regional rescue centers, and local rescue bases. A head rescue center is in charge of the
rescue operations at any demand point within the disaster area. Under the head rescue center, a regional res-
cue center is set up for each administrative district to control rescue operations at any demand point within the
district and to oer the backup to demand points in other districts. Local rescue bases in each district are
responsible for the rst-line rescue operation in their own administrative territory.
The division of the disaster area is usually based on administrative districts due to the consideration of com-
mand and authority. For the purpose of mutual backup between nearby districts, the disaster area is divided
into several groups according to the distance between districts, and each group consists of several administra-
tive districts. Districts in the same rescue group assist one another rst. In some situations, however, assistance
between districts in dierent rescue groups is also permitted.
Furthermore, disaster districts in the same group are classied into three levels that indicate the dierent
imperiousness of the rescue demand. The rst-level disaster district has the highest risk of being attacked
by a ood, while the third-level disaster district has the lowest risk. Therefore, rescue centers or bases in lower
level districts can backup higher level districts. That is, the backup of the head storehouse in the rst-level dis-
trict is the regional storehouse in the second-level district of the same group. The backups of the regional
storehouse in the second-level district are the local storehouse in the third-level district of the same group
and the regional storehouse in the second-level district of the other group.
It is evident that ood emergency logistics preparation planning is rigid and complex due to uncertain
demands. In this study, the ood emergency logistics preparation problem with uncertainty is formulated
as two stochastic programming models that determine the rescue resource distribution plan for urban ood
disasters, including the structure of rescue organization, the location of rescue resource storehouses, the allo-
cation of rescue resources within capacity restrictions, and the distribution of rescue resources. The basic
assumptions of this paper are described as follows:

Fig. 1. The structure of rescue organization and distribution.

 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754 741

1. The rescue resource may be supplied either by the government or by private enterprises with a rental
contract.
2. The stochastic rescue demands are predictable under all possible ooding scenarios.
3. Transportation costs are deterministic.

3.1. Notations and denitions

The following notations have used in the present mathematical model.

L(x) set of demand points under rainfall situation x
Lk(x) set of demand points that lie within the service area of regional rescue center kunder rainfall situation
x
Ln(x) set of demand points that belong to the Level-n areas under rainfall situation x
Lm n x set of demand points that belong to Group-m and Level-n areas under rainfall situation x
Jm n x set of local rescue bases that belong to Group-m and Level-n area under rainfall situation x
Km n x set of regional rescue centers that belong to Group-m and Level-n area under rainfall situation x
K 2 x set of regional rescue centers that belong to Level-2 areas under the rainfall situation x
M(l) group of demand point l
N(j) level of local rescue base j
Ci unit penalty for shortage of equipment type i during the period of rescue operation.
d il x amount of equipment type i desired at demand point l under rainfall situation x
ei average cost of equipment type i amortized to the period of rescue operation, including purchase cost
and maintenance cost
fj average operation cost of local rescue base j during the period of rescue operation
Oi unit pumpage of equipment type i
P(x) probability of rainfall situation x
Qj storage capacity of local rescue base j
Qk storage capacity of regional rescue center k
Qo storage capacity of head rescue center o
ri unit rent of equipment type i
tijl unit transportation cost of equipment type i shipped to demand point l from local rescue base j
tikl unit transportation cost of equipment type i shipped to demand point l from regional rescue center k
tiol unit transportation cost of equipment type i shipped to demand point l from head rescue center o
vi volume of equipment type i
dkl shortest distance from regional rescue center k to demand point l

All of the above terms dene the inputs of the model. The following decision variables are used in the
formulation.
qijl x amount of equipment type i shipped to demand point l from local rescue base j under rainfall situation
x
qikl x amount of equipment type i shipped to demand point l from regional rescue center k under rainfall
situation x
qiol x amount of equipment type i shipped to demand point l from head rescue center o under rainfall sit-
uation x
S ij amount of equipment type i stored in local rescue base j
S ik amount of equipment type i stored in regional rescue center k
S io amount of equipment type i stored in head rescue center o
X mnk if regional rescue center k belongs to group m and level n, this equals one; otherwise this equals zero
yj if local rescue base j are selected for setup, this equals one; otherwise, this equals zero
Ym k0 k if regional rescue centers k 0 and k both belong to group m, this equals one; otherwise, this equals zero
i
al x amount of shortage of equipment type i to demand point l under rainfall situation x
bij x amount of surplus of equipment type i in local rescue base j under rainfall situation x
742 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754

bik x amount of surplus of equipment type i in regional rescue center k under rainfall situation x
bio x amount of surplus of equipment type i in head rescue center o under rainfall situation x

3.2. The grouping and classifying model

This model aims to group the disaster rescue areas and classify their level of emergency by minimizing the
expected shipping distance. Disaster rescue areas are classied into several multi-level disaster rescue groups.
Thus, each rescue supplier belongs to a specic disaster rescue group and a level of emergency
2 3
XXX X X X
min 4Y m0 P x dk0 l d il x5 1
kk
m k k 0 6k x i l2Lk x
X X
Subject to 2Y mk0 k 6 X mn
k0 X mn m
k 6 Y k0 k 1 8m; k 0 ; k 2
n n
X X X X X X
P x Oi d il xX m1
k P P x Oi d il xX m2
k0 8m; k; k 0 6 k 3
x l2Lk x i x k0
l2L x i
X X X X X X
P x Oi d il xX m2
k P P x Oi d il xX m3
k0 8m; k; k 0 6 k 4
x i x 0 i
l2Lk x l2Lk x
X
X mn
k P 1 8m; n 5
k
XX
X mn
k 1 8k 6
m n

Y mk0 k 2 f0; 1g 8m; k; k 0 7

X mn
k 2 f0; 1g 8m; n; k 8
The objective of the model, Eq. (1), is to minimize the expected shipping distance of rescue equipment. Eq.
(2) states that Y mk0 k equals one if both regional rescue centers k and k 0 belong to group m; otherwise Y mk0 k equals
zero. Eq. (3) requires that the expected amount of ood drained at the demand points, which belong to the
rst-level regional rescue center k, must be greater than or equal to that of any second-level regionals rescue
center k 0 in the same group. Eq. (4) represents that the expected amount of water drained at the demand points
in the second-level rescue region must be greater than or equal to that of any third-level rescue region in the
same group m. Eq. (5) states that every level in each group should contain at least one regional rescue center.
Eq. (6) restricts that a regional rescue center can only be classied to a certain level of a specic group. Eqs. (7)
and (8) dene Y mk0 k and X mn k as binary variables.
For short, the objective function of this model can be rewritten as
min EC 1 X; Y; dn 10
X;Y2X1

3.3. The location-allocation model

Based on the results of the above model, the following two-stage stochastic programming model determines
the selected local rescue bases that need to be set up after the disaster as well as the quantity of rescue equip-
ment in the storehouses of all echelons and the transportation plans of rescue equipment. The former is the
rst-stage decision and the two latter cases are the second-stage decisions based upon the rst-stage decision
and the realized uncertain scenarios. The corresponding costs for distributing the rescue equipment in all res-
cue demand scenarios must be considered in the rst-stage decision. Thus, the objective is to minimize the cur-
rent facility setup cost and the equipment average cost as well as the expected future transportation cost, the
supply-shortage cost, and the demand-shortage penalty during the period of rescue operation. The mathemat-
ical formulation of this model is as follows:
 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754 743
" #
X X X X X X X
min fj y j e i
S ij S ik S io P xTCx P xRCx P xSCx
j i j k x x x

9
2 3
X X6 X X X 7
Subject to TCx 4 tijl qijl x tikl qikl x tiol qiol x tikl qikl x5
l2L1 x i j2J 1
M l
k2K 1
M l Ml
k2K 2
2 3
X X6 X X X 7
4 tijl qijl x tikl qikl x tiol qiol x tikl qikl x5
l2L2 x i Ml
j2J 2 k2K 2 M l
k2K 3
2 3
X X6 X X 7
4 tijl qijl x tikl qikl x tiol qiol x5 10
l2L3 x i Ml
j2J 3
Ml
k2K 3
X X
RCx ri ail x 11
i l
" #
X X X
SCx C i
bij x bio x bik x 12
i j k
X X X
ail x d il x  qijl x  qikl x  qiol x  qikl x 8i; l 2 L1 x; x 13
Ml Ml Ml
j2J 1 k2K 1 k2K 2
X X X
ail x d il x  qijl x  qikl x  qiol x  qikl x 8i; l 2 L2 x; x 14
Ml
j2J 2 k2K 2 M l
k2K 3
X X
ail x d il x  qijl x  qikl x  qiol x 8i; l 2 L3 x; x 15
Ml Ml
j2J 3 k2K 3
X
bij x S ij  qijl x 8i; j; x 16
Mj
l2LNj x
X
bik x S ik  qikl x 8i; k 2 K 1 ; x 17
Mk
l2L1 x
X X
bik x S ik  qikl x  qikl x 8i; k 2 K 2 ; x 18
Mk
l2L1 x l2L2 x
X X
bik x S ik  qikl x  qikl x 8i; k 2 K 3 ; x 19
Mk M k
l2L3 x l2L2 x
X
bio x S io  qiol x 8i; x 20
l2Lx
X
vi S ij 6 Qj y j 8j 21
i
X
vi S ik 6 Qk 8k 22
i
X
vi S io 6 Qo 23
i

y j 2 f0; 1g 8j 24
For all S ij P 0; S ik P 0; S io P0 25
For all qijl x P 0; qikl x P 0; qiol x P 0 26
For all ail x P 0; bij x P 0; bik x P 0; bio x P0 27
744 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754

The rst term of Eq. (9), the sum of facility setup costs of local rescue bases fj, includes the necessary costs
to set up local rescue bases during the period of rescue operation, such as expenses for water, electricity, and
personnel. The second term of Eq. (9) denes the average costs for all rescue equipment.
Under rainfall situation x, Eq. (10) denes the transportation cost of all rescue equipment, Eq. (11) shows
the rental cost for the insucient equipment, and Eq. (12) denes the penalties of surplus equipment. Herein,
the equipment rent is dened as the supply-shortage charge, and the equipment average cost is dened as the
demand-shortage penalty.
Eqs. (13)(15) dene the amount of shortage of equipment type i at demand point l under rainfall situation
x, which is represented by the dierence of the demand d il x and the inow of equipment to the demand
point l. For example, Eq. (13) states that the equipment required at the rst-level demand point comes from
the rst-level local rescue bases, rst-level and second-level regional rescue centers of the same group, and the
head rescue center. The shortage is designated as ail x, which needs to be rented from a private enterprise
(i.e., contractor).
Eqs. (16)(20) dene the amount of surplus equipment type i under the rainfall situation x in each store-
house, which is represented by the dierence in the storage and the outow of equipment.
Eqs. (21)(23) show the capacity constraints of storehouses. Eq. (24) denes yj as one if the local rescue base
j is set up; otherwise yj equals to zero. Eqs. (25)(27) are non-negative constraints. For short, the objective
function of this model can be rewritten as
" #
X X X X
i i i i
min fj y j e Sj S k S o EC 2 y; S; dn 90
y;S;q
j i j k
a;b2X2

4. Solution procedure

In this section, we will develop a solution procedure for solving the ood emergency logistics preparation
problems, Eqs. (1)(27). As previously mentioned, a key diculty in solving the stochastic program is evalu-
ating the expectation in the objective; it is dicult to estimate the rescue equipment demand for dierent rain-
fall situations. There are far too many ooding scenarios for the existing stochastic programming approaches
to handle. The goal of scenario planning is to specify a set of scenarios that represent the possible realizations
of unknown problem parameters and to consider the range of scenarios in determining a robust solution.
Thus, we can examine these problems by using the sample average approximation (SAA) scheme. The detailed
steps of SAA are given in Section 4.2.
Flooding potential maps at dierent rainfall situations are used for developing scenarios that present all
possible locations of rescue demand points and the amount of rescue equipment. Applying the data processing
and network analysis functions of the geographic information system (GIS), the detailed derivation of rescue
demand point locations and rescue equipment demands for dierent ooding scenarios is shown in Section
4.1.

4.1. Scenario development: applying geographic information system (GIS) technique

The locations of demand points and the quantity of required rescue equipment under dierent rainfall sit-
uations can be estimated through the GIS analysis function. Then, the estimated corresponding quantity of
required rescue equipment can be used to determine the quantity of required rescue equipment for demand
points under dierent ooding scenarios. The data processing procedure is explained in the following.

Step 1. Generating map layers

Step 1.1: Generate a point map layer that indicates the locations of a head rescue center, regional rescue cen-
ters, and candidates for local rescue bases. The information about the storage capacity and the oper-
ation cost of rescue centers and bases is stored in the database of the point map layer.
Step 1.2: Generate polyline map layers, such as road map layers. The information about the width of roads is
stored in the database of the polyline map layers.
 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754 745

Step 1.3: Generate polygon map layers, including an administrative district map layer, a sub-district map
layer, and ooding potential map layers. The information about the depth of oodwater is stored
in the database of the polygon map layers.
Step 2. Overlapping analysis of map layers
Step 2.1: Overlap the ooding potential map layers with the administrative sub-district map layer. Apply the
intersecting operation to generate the map layers of ooded sub-districts under dierent rainfall
situations.
Step 2.2: According to the map layers of ooded sub-districts, nd the centroid of each ooded spot to rep-
resent the rescue demand point. Apply buering and intersecting operations to generate the demand
point map layers under dierent rainfall situations.
Step 2.3: Overlap supply point map layer and demand point map layers to generate the network point map
layers under dierent rainfall situations.
Step 3. Data analysis and output
Step 3.1: According to the map layers of ooded sub-districts, calculate the volume of oodwater at each
ooded spot, and estimate the required quantities of each kind of rescue equipment at each demand
point under dierent rainfall situations.
Step 3.2: Based on the results of Step 3.1, the quantity of required rescue equipment for each demand point
under dierent ooding scenarios is outputted.
Step 3.3: Establish the database of each kind of rescue equipment, including the purchase cost, the rental cost,
the transportation cost, etc.
Step 3.4: Output the origindestination distance matrices, which is derived from the shortest-path analysis by
using the information in the network point map layers and the road map layer.

4.2. Sample average approximation

Since the exact evaluation is dicult or even impossible, the expected objective value is often approximated
through sample averaging. Sampling-based approximation may consist of internal or external methods. Inter-
nal sampling methods (e.g. stochastic decomposition algorithm, and stochastic branch and bound algorithm)
perform sampling inside an algorithm with new i.i.d. samples generated and accumulated over iterations; the
entire history of samples is used at each iteration.
External sampling, which is also called sample average approximation, approximates the true problem by
the sample average approximation problem. Hence, once the sample is generated, the existing deterministic
algorithms can solve the stochastic problem. Kleywegt et al. (2001) show that the SAA method converges
to the optimal solution and solution value. Two proposed stochastic programming models will be solved using
the SAA method. The detailed steps of the solution algorithm are as follows:
Step 1. Generate V independent samples of ooding scenarios each of size U, i.e. fd1v n; . . . ; dUv ng;
8v 1; . . . ; V , where d fd il x; 8i; j; xg. For each sample, solve these two corresponding problems sequen-
tially. Let ^zv1 and X b m; Y
b v ; m 1; . . . ; V , be the corresponding optimal objective value and optimal solution of
v
model 1, and let ^z2 and ^ b v ; m 1; . . . ; V , be the counterparts of model 2.
ym ; S
1 X
Model 1 : min EC 1 X; Y; duv n 28
X;Y2X1 U
u
" #
X X X X 1 X
Model 2 : min fj y j ei i
Sj i i
Sk So EC 2 y; S; duv n 29
y;S;q;
j i j k
U u
a;b2X2

Step 2. Compute zi and r2zi using the following Equations.

It is well known that zi provides a lower statistical bound for the optimal objective value zi and that r2zi is an
estimate of the variance of this estimator for these two models.
1 X v
zi : ^z ; i 1; 2 30
V v i
746 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754

1 X
r2zi : ^zv  zi 2 ; i 1; 2 31
V  1V m i

Step 3. Choose two feasible solutions X; e Y

e and ~y; Se from the above computed solutions X
b m; Y
b v and
^m b v
y ; S . Estimate the true objective function value as
X
~z1 X; e : 1
e Y e Y;
EC 1 X; e du n 32
0
U u
" #
X X X X 1 X
min ~z2 ~ e :
y; S fj ~y j ei e
S ij S ik e
e S io 0 e du n
EC 2 ~y; S; 33
q;a;b2X2
j i j k
U u
0
Generate another U 0 independent samples of the ooding scenarios, i.e., fd1 n; . . . ; dU ng; 8u 1; . . . ; U 0 .
Estimate the variances of these estimators ~z1 X; e Y e as
e and ~z2 ~y; S
1 Xh i2
e Y
r~2z1 X; e : EC 1 e Y;
X; e du n  ~z1 X;
e Y
e 34
U 0  1U 0 u
( " # )2
1 X X X X X 1 X
r~2z2 ~ e :
y; S fj ~y j ei e
S ij S ik e
e S io 0 e du n  ~z2 ~y; S
EC 2 ~y; S; e
U 0  1U 0 u j i j k
U u
35

Typically, U 0 is much larger than the sample size U used in solving the SAA problems.
Step 4. Compute the estimate of the optimality gap of the solutions X;e Y e respectively.
e and ~y; S,
e Y
gap1 X; e : ~z1 X;
e Y
e  z1 36
gap ~ e : ~z2 ~
y; S e  z2
y; S 37
2

The estimated variances of the above gap estimators are then given by
e Y
r2gap1 r~2z1 X; e r2 38
z1

r2gap2 r~2z2 ~
y; e r2
S 39
z2

5. Computational results

This section describes the ndings of the model implementation. We will describe the characteristics of the
test problem, implementation details, and demonstrate the sensitivity analysis.

5.1. Data and implementation

The proposed model is applied to Taipei City, located in Northern Taiwan. Taipei City is 272 square kilo-
meters wide and has a population density of over 10,000 people per square kilometer. Many places in Taipei
City often face the risk of ood during typhoon seasons. To reduce the damage of disasters, the government
requires that the entire disaster rescue operation should be completed within 36 hours. Furthermore, the Tai-
pei City Government is designated as the head disaster rescue center; 12 administrative district oces are set
up as regional rescue centers; and 84 oces of the subdivision of the district are chosen as candidate points for
the local rescue bases.
The allocation of rescue resources is an important issue in ood rescue operation management. Small-size
auto prime engine pumps are among the most important rescue equipment often used in areas with troubled
water drainages. For the sake of demonstration, we will focus on these pumps in the following example. There
are two kinds of small-size auto prime engine pumps, i.e. Type A and Type B. They dier in pumpage and
have no signicant dierence in their sizes and unit transportation costs. The pumping power of Type B is
 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754 747

Table 1
Parameters setting of rescue equipment
Equipment Unit rent of equipment Average operating cost of Unit penalty for shortage of Unit pumpage
type (NT$) equipment (NT$) equipment (NT$) (m3/h)
Type A 35640 67 67 72
Type B 36396 94 94 108

Table 2
Parameters setting of rescue facility
Facility type Storage capacity (pumps) Average operating cost of facility (NT$)
Head rescue center 281
Regional rescue center 35
Local rescue base 22 382

higher than Type A. The transportation cost of pumps is included in the rental fee, which is assumed identical
in dierent rescue areas.
The proposed model for the ood emergency logistics preparation requires data for various parameters (see
Tables 1 and 2). The location of rescue demand points and the required rescue equipment are determined in
advance according to the ooding potential maps under dierent rainfall situations. The rainfall is classied
into three situations: 300 mm, 450 mm and 600 mm per day, with the probabilities 0.208, 0.025 and 0.003
respectively. The ood tendency graphs are vectorized and overlaid with the administrative district map
(see Fig. 2).

5.2. Results of ood emergency logistics preparation planning

The GIS package ESRI ArcGIS 9.x was used to perform the spatial data analysis of the testing problem.
The alternative locations of the rescue centers/bases and an origindestination matrix are shown in Figs. 3 and
4, respectively. The related results of the oodwater estimation are displayed in Fig. 5. The number of rescue
demand points is 35, 91, and 159 at the three rainfall situations, respectively.

Fig. 2. Vectorizing picture of ooding potential map (precipitation: 600 mm/day).

748 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754

Fig. 3. Alternative location of rescue center/base.

Fig. 4. Origindestination distance matrix.

After the spatial data analysis, the mathematical programming package LINGO was used to solve the pro-
posed models. To implement the SAA method, we set the size of ooding scenarios V, the number of small-
size samples U, and the number of big-size samples U 0 as 3, 30, and 2000, respectively. The 12 administrative
districts in Taipei City are divided into four disaster rescue groups, each classied into three levels. The results
of the rescue organization structure are indicated in Table 3. In each disaster rescue group, the distribution of
rescue demand points is displayed in Figs. 68. The distribution of all disaster rescue centers/bases is illus-
trated in Fig. 9, and the storage amounts in each disaster rescue supply point are presented in Table 4. Fur-
thermore, the total storage amounts of the Type A pump and the Type B pump are 642 and 232, respectively.
Note that only one head rescue center, 10 regional rescue centers, and 15 local rescue bases are sucient for
Taipei City to perform the ood rescue operations.
 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754 749

Fig. 5. Distribution of ooding.

Table 3
Planning result of the rescue organization structure
Rescue groups Rescue level
Level 1 Level 2 Level 3
Group 1 10 9 1
Group 2 4 12 11
Group 3 8 3 2
Group 4 7 5 6

Fig. 6. Estimated result of ooded point distribution (precipitation: 300 mm/day).

750 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754

Fig. 7. Estimated result of ooded point distribution (precipitation: 450 mm/day).

Fig. 8. Estimated result of ooded point distribution (precipitation: 600 mm/day).

5.3. Quality of stochastic solutions

In this section, we will compare the real ood emergency logistics system (Y0, X0, y0, S0) with the solution
of the stochastic programming models and the solution of the deterministic optimization model involving
the mean value of uncertain demands. The latter is named by the mean-value problem and denoted as
 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754 751

Fig. 9. Planning results of rescue center/base location.

Table 4
Result of the storage amount of equipment
Equipment type Facility type
Head rescue center Regional rescue center
1 1 2 3 4 5 6 7 8 9 10 11 12
Type A 264 9 0 33 26 30 0 25 26 13 13 2 28
Type B 17 1 0 2 9 5 0 10 9 22 22 7 7

Equipment type Local rescue base

22 23 24 25 26 45 47 48 50 51 58 59 60 61 62
Type A 20 21 21 22 21 0 4 14 1 18 10 1 1 0 19
Type B 2 1 1 0 1 22 18 5 21 4 12 11 1 20 2

Table 5
Cost statistics for candidate solutions
Total rescue cost Candidate solution
(Y0, X0, y0, S0) (YM, XM, yM, SM) (Y1, X1, y1, S1)
Average 325501.54 230231.22 228740.94
Std. dev. 207.78 187.72 122.11
Min. 318145.80 226353.10 225500.20
Max. 332126.70 236714.70 231557.70
Gap 3707.57 2438.12 353.28
rgap 382.52 252.76 228.17
Note: Gap: Optimality gap, as dened in Eq. (37) rgap: standard deviation of the gap estimate, as dened in Eq. (39).

(YM, XM, yM, SM). The best candidate solution obtained from solving the SAA problems is noted as
(Y1, X1, y1, S1). The total rescue cost statistics for each solution are computed with a sample size of
752 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754

U 0 = 2000, as shown in Table 5. It is obvious that the solution of the stochastic programming model is best in
terms of average cost, optimality gap, and variability of the total cost. This candidate solution (Y1, X1, y1, S1)
is not only superior to the traditional mean-value problem solution in the sense of expectation but also results
in smaller total cost variability.

5.4. Sensitivity analysis

The storage capacity at each rescue echelon was adjusted proportionally to reveal the impact on total rescue
cost. Taking the adjustment of the storage capacity at the head rescue center as an example, the storage capac-
ity is set as 422 when the adjustment ratio is 1.5. Other storage capacity at each rescue echelon is unchanged.
The total storage capacity is 2694 pumps. The results of single parameter sensitivity analysis of the storage
capacity of the head rescue center, regional rescue centers, and local rescue bases are shown in Fig. 10.

310000.0

290000.0
Average total rescue cost

270000.0

250000.0

230000.0

210000.0

190000.0

170000.0
0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0
Adjustment ratio of storage capacity

head rescue center regional rescue center local rescue base

Fig. 10. Results of sensitivity analysis about facility capacity.

25000.00

20000.00
Total Rescue Cost Range

15000.00

10000.00

5000.00

0.00
Low Medium High
Variability
SAA MVP

Fig. 11. Eect of variability on the total rescue cost.

 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754 753

The following conclusions about the objective value of the location-allocation model were drawn from the
results of the sensitivity analysis of storage capacity. As the storage capacity at each rescue echelon increases,
the average total rescue cost reduces and vice versa. Under the same adjustment ratio of storage capacity, the
increase in storage capacity at the regional rescue centers conduces to a maximal drop-o of the average total
rescue cost. For example, the increase in storage capacity at the head rescue center, regional rescue centers,
and local rescue bases correspondingly causes the average total rescue cost to drop to 227725.00,
197887.94, and 224617.40 (NT$) when the adjustment ratio of storage capacity is 1.5. These results conrm
that allocating the correct amount of rescue resource to the proper place can lead to the best outcome.
In order to observe the eect of the variability of the uncertain demand, three dierent levels of the vari-
ability of the uncertain demand were considered. In Fig. 11, the range of the total rescue cost for the best can-
didate solution is compared with the corresponding mean-value problem solution. As mentioned, the
stochastic solution has smaller ranges than the mean-value problem solution. Furthermore, with the increase
in the variability of the uncertain demand, the range of total rescue cost of the stochastic solution increases at
a slower rate than that of the mean-value problem solution.

6. Conclusions and future research

In many countries, a ood rescue commanding system is elaborately constituted for the sake of mutual
assistance of rescue organizations. Mutual assistance agreements exist among administrative districts and
between the government and private groups. One practical feature of ood emergency logistics is a multi-
group, multi-echelon, and multi-level rescue network. In this paper, we have developed a methodology for
ood emergency logistics preparation planning under uncertainty. Two stochastic programming models were
proposed to assist the government agents in determining the rescue resource storehouse locations, required
quantities of rescue equipment, and the distributions of the rescue equipment.
Improving the eciency of emergency disaster relief must depend on circumspect preparation plans made
in peacetime. Because natural disasters occur frequently in many countries around the world, governments
invest in disaster management system to reduce loss of life and property. Most of these systems follow the
GIS framework. Recently, many disaster-relevant databases have been established with the assistance of
GIS graphical presentations and manmachine interfaces. Such applications include analyzing transportation
lifelines, identifying areas with high degrees of damage, and classifying the available resources. These GIS-
based disaster management systems usually do not provide any specic analysis modules for planning
emergency logistics. In this study, however, by using the data obtained from the GIS-based disaster prediction
subsystem and rescue resource subsystem, an optimal ood emergency logistics preparation plan can be deter-
mined. Numerical tests reveal the computational eectiveness of this model. Such a ood emergency logistics
system has a better capability to combine with other disaster management operation systems and to give
adequate decision support for government agencies in both theory and practice.
Finally, we make the following related suggestions for further research:

1. For a more practical application, the proposed methodology can be developed into a plug-in software pack-
age with a friendly graphical user interface. This module may be integrated into the GIS-based disaster
management system in the future. In addition, prior ood data should be added to the demand database
in order to match the demand situation in reality, although ooding potential maps can predict areas that
are more likely to ood. Further developments in on-line GIS-based disaster management systems are
needed.
2. Although we have not addressed the vehicle scheduling problem, the restraints of rescue vehicle types and
the limitations of rescue manpower should be considered in ood emergency logistics management.
3. In this research, we have considered the ood emergency logistics problem for a single urban city. How-
ever, our model can be extended to examine the rescue resources sharing problems among multiple urban
cities.
4. Meta-heuristics have been applied successfully to the deterministic optimization program, but only a few
meta-heuristics have been developed for their stochastic optimization program. Developing a meta-heuris-
tic would be a particularly valuable approach to solve realistic ood emergency logistics problems.
754 M.-S. Chang et al. / Transportation Research Part E 43 (2007) 737754

Acknowledgement

This work was supported in part by the National Science Council, Taiwan, under Grant No. NSC 95-2416-
H-033-013.

References

Aikens, C.H., 1985. Facility location models for distribution planning. European Journal of Operational Research 22 (3), 263279.
Fiedich, F., Gehbauer, F., Rickers, U., 2000. Optimized resource allocation for emergency response after earthquake disasters. Safety
Science 35, 4157.
Georion, A.M., Powers, R.F., 1995. Twenty years of strategic distribution system design: an evolutionary perspective. Interfaces 25, 105
128.
Gutierrez, G.J., Kouvelis, P., Kurawala, A.A., 1996. A robustness approach to uncapacitated network design problems. European Journal
of Operations Research 94, 362376.
Haghani, A., Oh, S., 1996. Formulation and solution of a multi-commodity, multi-modal network ow model for disaster relief
operations. Transportation Research A 30, 231250.
Kleywegt, A.J., Shapiro, A., Homen-De-Mello, T., 2001. The sample average approximation method for stochastic discrete optimization.
SIAM Journal of Optimization 12, 479502.
Knott, R., 1987. The logistics of bulk relief suppliers. Disaster 11, 113115.
Knott, R., 1988. Vehicle routing for emergency relief management: a knowledge-based approach. Disaster 12, 285293.
MirHassani, S.A., Lucas, C., Mitra, G., Messina, E., Poojari, C.A., 2000. Computational solution of capacity planning models under
uncertainty. Parallel Computing 26, 511538.
Owen, S.H., Daskin, M.S., 1998. Strategic facility location: a review. European Journal of Operational Research 111, 423447.
zdamar, L., Ekinci, E., Kucukyazaci, B., 2004. Emergency logistics planning in natural disasters. Annals of Operations Research 129,
O
217245.
Ray, J., 1987. A multi-period linear programming model for optimally scheduling the distribution of food-aid in West Africa. M.S. Thesis,
University if Tennessee, Knoxville, TN.
Santoso, T., Ahmed, S., Goetschalckx, M., Shapiro, A., 2005. A stochastic programming approach for supply chain network design under
uncertainty. European Journal of Operational Research 167, 96115.
Tsiakis, P., Shah, N., Pantelides, C.C., 2001. Design of multiechelon supply chain networks under demand uncertainty. Industrial and
Engineering Chemistry Research 40, 35853604.
Vidal, C., Goetschalckx, M., 1997. Strategic productiondistribution models: a critical review with emphasis on global supply chain
models. European Journal of Operational Research 98, 118.
Vidal, C., Goetschalckx, M., 2001. A global supply chain model with transfer pricing and transportation cost allocation. European
Journal of Operational Research 129, 134158.