2 Scopusresults

Geoscience Frontiers 12 (2021) 101095
Contents lists available at ScienceDirect
Geoscience Frontiers
journal homepage: www.elsevier.com/locate/gsf
Research Paper
Location-allocation modeling for emergency evacuation planning with

GIS and remote sensing: A case study of Northeast Bangladesh
Mahfuzur Rahman a,b,c, Ningsheng Chen a,b,⁎, Md Monirul Islam c, Ashraf Dewan d,
Hamid Reza Pourghasemi e, Rana Muhammad Ali Washakh f,g, Nirdesh Nepal h,
Shufeng Tian a,b, Hamid Faiz a,b, Mehtab Alam a,b, Naveed Ahmed b
a
Key Laboratory for Mountain Hazards and Earth Surface Process, Institute of Mountain Hazards and Environment (IMHE), Chinese Academy of Sciences (CAS), Chengdu 610041, China
b
University of Chinese Academy of Sciences (UCAS), Beijing 100049, China
c
Department of Civil Engineering, International University of Business Agriculture and Technology (IUBAT), Dhaka 1230, Bangladesh
d
School of Earth and Planetary Sciences, Curtin University, Kent St, Bentley, WA 6102, Australia
e
Department of Natural Resources and Environmental Engineering, College of Agriculture, Shiraz University, Shiraz, Iran
f
School of Architecture, Neijiang Normal University, Neijiang 641100, China
g
Institute of Tibetan Plateau Research, Chinese Academy of Sciences, Beijing 100101, China
h
Global Institute for Interdisciplinary Studies, Kathmandu 3084, Nepal
a r t i c l e i n f o a b s t r a c t
Article history: This work developed models to identify optimal spatial distribution of emergency evacuation centers (EECs) such
Received 11 May 2020 as schools, colleges, hospitals, and fire stations to improve flood emergency planning in the Sylhet region of
Received in revised form 27 August 2020 northeastern Bangladesh. The use of location-allocation models (LAMs) for evacuation in regard to flood victims
Accepted 29 September 2020
is essential to minimize disaster risk. In the first step, flood susceptibility maps were developed using machine
Available online 7 November 2020
learning models (MLMs), including: Levenberg–Marquardt back propagation (LM-BP) neural network and deci-
Handling Editor: Sanghoon Kwon sion trees (DT) and multi-criteria decision making (MCDM) method. Performance of the MLMs and MCDM tech-
niques were assessed considering the area under the receiver operating characteristic (AUROC) curve.
Keywords: Mathematical approaches in a geographic information system (GIS) for four well-known LAM problems affecting
Natural disasters emergency rescue time are proposed: maximal covering location problem (MCLP), the maximize attendance
Emergency evacuation centers (MA), p-median problem (PMP), and the location set covering problem (LSCP). The results showed that existing
Flooding EECs were not optimally distributed, and that some areas were not adequately served by EECs (i.e., not all de-
Machine learning mand points could be reached within a 60-min travel time). We concluded that the proposed models can be
Multi-criteria decision making
used to improve planning of the distribution of EECs, and that application of the models could contribute to
Location-allocation model
reducing human casualties, property losses, and improve emergency operation.
© 2021 China University of Geosciences (Beijing) and Peking University. Production and hosting by Elsevier B.V. This
is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
1. Introduction elevation. The flooding has severe implications on livelihoods, food

security, and loss of property (Armah et al., 2010). Worldwide estimates
Bangladesh has been affected by numerous devastating floods show that floods between 1995 and 2015 affected 2.3 billion people and
because of its geographic setting, heavy monsoon rainfall, and low led to 157,000 deaths (Wahlstrom and Guha-Sapir, 2015). Since the end
of the twentieth century, estimated global economic loss resulting from
Abbreviations: AUROC, area under the receiver operating characteristic; DEM, digital floods is huge (US $386 billion) (Wang et al., 2011). Modeling studies
elevation model; DT, decision trees; EEC, emergency evacuation center; ESWARA, have suggested that anthropogenic climate change could lead to more
Extended stepwise weight assessment ratio analysis; FR, frequency ratio; GIS, frequent and devastating floods (Hirabayashi and Kanae, 2009;
geographic information system; IG, information gain; LAM, location-allocation model;
Whitfield, 2012). To reduce the risk associated with floods and success-
LM-BP, Levenberg–Marquardt back propagation; LSCP, location set covering problem;
LC, land cover; MA, maximize attendance; MCDM, multi-criteria decision making; MCLP, ful management of this phenomenon, emergency evacuation planning
maximal covering location problem; MCT, multi-collinearity test; MFI, Modified Fournier can be of immense assistance (Kar and Hodgson, 2008; Hunt and
Index; ML, machine learning; MLM, machine learning model; NDVI, normalized difference Watkiss, 2011).
vegetation index; nFR, normalized frequency ratio; PMP, p-median problem; SPI, stream While the occurrence of floods cannot be avoided, damage associ-
power index; SWARA, stepwise weight assessment ratio analysis; TOL, tolerance; TWI, to-
pographic wetness index; VIF, variance inflation factor.
ated with it can be minimized substantially if flood susceptible areas
⁎ Corresponding author. can be delineated accurately. Hence, flood prediction is an important
E-mail address: chennsh@imde.ac.cn (N. Chen). step in reducing flood damage. In Bangladesh, the study of flood risk
https://doi.org/10.1016/j.gsf.2020.09.022
1674-9871/© 2021 China University of Geosciences (Beijing) and Peking University. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
M. Rahman, N. Chen, M.M. Islam et al. Geoscience Frontiers 12 (2021) 101095
management and emergency evacuation is at an early stage. The coun- In Bangladesh, many governmental, non-governmental, international,
try lacks a comprehensive emergency evacuation plan, particularly in private, and volunteer organizations are involved with the emergency
areas that are persistently affected by floods (Masuya et al., 2015). Sev- functions like dissemination of flood warning, evacuation, and relief op-
eral studies have been conducted to develop flood hazard maps to eval- eration among the flood victims. However, in most emergency rehabil-
uate flood risk in Bangladesh (Chowdhmy and Karim, 1996; Islam and itation programs, the operational team is unable to reach affected areas
Sado, 2000b, 2002; Tingsanchali and Karim, 2005; Dewan et al., 2007; in due time because of lack of appropriate information. Moreover, the
Islam, 2011; Masood and Takeuchi, 2012; Hossain, 2013; Ganguly location of the EECs is generally unknown and emergency management
et al., 2019). Even though flood protection works have been undertaken, finds it very difficult to transfer potential flood victims to appropriate
majority of them are conducted locally and without pre-planning or a EECs during floods, especially in the areas having lack of flood control
detailed flood risk assessment. It is, therefore, essential to consider (Alam and Ali, 2002). It is also difficult to determine whether these
flood susceptibility mapping and risk assessment, and to develop emer- EECs are operational, because some cannot be placed optimally (Kar
gency evacuation plans for pre- and post-flood situations so that the de- and Hodgson, 2008). As an evacuation plan depends on understanding
gree of damage can be reduced (Khosravi et al., 2016; Haghizadeh et al., the spatial distribution of EECs (Sherali et al., 1991), optimal allocation
2017). Though there have been several studies assessing flood suscepti- of EECs in susceptible areas is important. To optimize allocation of
bility of Bangladesh based on satellite images and hydrodynamic EECs, this study proposes four LAMs: i.e., p-median problem (PMP),
modeling (as well as other techniques and tools), none of the studies maximal covering location problem (MCLP), location set covering prob-
have integrated location-allocation models (LAMs), multi-criteria deci- lem (LSCP), and maximize attendance (MA). The results of the LAMs can
sion making (MCDM) technique, and machine learning (ML) algorithms provide scientific evidence for planning optimal locations for a variety of
to develop a comprehensive evacuation plan. The current study intends LAM users. Furthermore, location-allocation study along with ML and
to fill this gap. MCDM approaches would be innovative and makes new contributions
Although geospatial techniques have shown immense potential for to the management of flood risk. Specifically, the following contribu-
evacuation planning, their use remains limited in the context of tions to the science and society are explicitly noted. First, we attempted
Bangladesh floods. Geospatial techniques in ML can be used to construct to concentrate on synchronizing preparatory and response processes by
high quality datasets for machine learning models (MLMs) (Bui et al., integrating two approaches: i.e., the issue of locating facilities to identify
2019). Some of the recently proposed techniques for integration of hy- new EECs and the challenge of service allocation for the flood victims.
draulic, statistical, ML, and MCDM methods along with GIS and RS for Therefore, we can ascertain that a combined pre- and post-flood risk re-
flood susceptibility mapping include: FLO-2D (Silva, 2016; Erena et al., duction involves an integration of these issues. Secondly, we noted that
2018), HEC-RAS (Anees et al., 2016; Haltas et al., 2016), MIKE 21 designated evacuation centers should be given priority for restoration
(Hossain, 2013), frequency ratio (FR: Samanta et al., 2018; Ali et al., and expansion to achieve the integration of resources to solve a problem
2019), logistic regression (LR: Al-Juaidi et al., 2018), weights-of- relevant to LA.
evidence (WoE: Hong et al., 2018; Tehrany et al., 2019), artificial neural In this study, we hypothesized that the ML and MCDM models per-
networks (ANNs: Wahab and Ludin, 2018; Falah et al., 2019), decision form better in predicting flood and the results may be useful for identi-
trees (DT: Lee et al., 2017; Janizadeh et al., 2019), support vector ma- fying emergency evacuation centers (EECs) using LAMs. Specific study
chine (SVM: Chang et al., 2018; Bermúdez et al., 2019), analytic hierar- aims are to: (i) develop a flood susceptibility map using ML and
chy process (AHP: Seejata et al., 2018; Cabrera and Lee, 2019), simple MCDM approaches, incorporated with geospatial techniques; (ii) assess
additive weighting (SAW: Jato-Espino et al., 2019), and Technique for efficiency of ML and MCDM on the spatial distribution and predictive
Order Preference by Similarity to Ideal Solution (TOPSIS: Khosravi ability of flood susceptible areas; (iii) compare two ML techniques
et al., 2019; Kim et al., 2019). Furthermore, in recent years some ensem- (Levenberg–Marquardt back propagation (LM-BP) neural network and
ble models have been proposed to improve the efficiency of ML tech- decision trees (DT)) with the MCDM method to develop a population
niques (Tehrany et al., 2015; Chapi et al., 2017). risk map; (iv) assess the spatial distribution of flood EECs and quantify
In general, ML and MCDM models cannot cover multidimensional the number of EECs to serve the population demand points
problems, such as allocating the location of emergency evacuation cen- i.e., children, disabled and elderly in different impedance cutoffs using
ters (EECs) for flood victims. However, they can be used to evaluate four LAMs; and (v) develop a PMP based optimized MCLP model to
flood probability in an affected region. The integration of ML, MCDM serve the maximum demand points.
and LAMs in GIS would overcome this issue and optimally allocate
EECs to appropriate locations in flood-prone regions for serving the 2. Materials and methods
flood victims. In addition, an integrated approach is not only limited
to flood evacuation planning (pre- and post-flood) in Bangladesh but 2.1. Study area
also could be applied to similar environments. Besides, this approach
is feasible to apply in managing different disasters i.e., earthquake, land- Sylhet is an area subject to persistent flooding and lies in northeast-
slide, and cyclone, etc. Moreover, to the best of our knowledge, this is ern Bangladesh. The area is surrounded by the Meghalaya hill to the
the first-hand work to suggest optimization process of large-scale emer- north (approximately 25.47°N and 91.37°E), the Tripura and Mizoram
gency planning, in the context of disaster preparedness and recovery re- ranges to the south (approximately 23.16°N to 23.94°N and 91.98°E to
sponse. However, LAMs have rarely been applied to planning of flood 92.94°E), and the highlands of Manipur to the east (approximately
evacuation, especially in Bangladesh. A few studies have investigated in- 24.66°N and 93.91°E). It has a population of more than 12.1 million peo-
corporation of geospatial techniques into LAMs in different countries ple (BBS, 2019) with an area of 12,298.4 km2 (population density of
(Chang et al., 2007; Gu et al., 2010; Li et al., 2011; Romero et al., 2012; about 980 persons per km2) (Fig. 1). The major rivers are Patnai, Baulai,
Xi et al., 2013; Mestre et al., 2015; Polo et al., 2015; Dinu and Ciucur, Jadukata, and Chalti in the north which regulate flow into Bangladesh
2016; Khan et al., 2018; Xu et al., 2018; Rizeei et al., 2019). from India (CEGIS, 2012). Hilly regions with steep slopes collect rainwa-
To support effective flood management planning, the current study ter as discharge and inundate river plains of the northeast region. The
attempts to develop an evacuation system. Role of existing EECs, estab- physiography includes hills and a few large depressions locally known
lishment of temporary shelters, support for rescue and health care facil- as ‘Beels’, which may be categorized mainly as oxbow lakes. Geologi-
ities, and supply of food, medication, and sanitary kits are necessary cally, the area is complex, with different geomorphic features. It has a
activities during a flood. Furthermore, it is essential to provide emer- high topography from the Plio-Miocene period, including the Khasi
gency services in a timely manner in order to reduce disaster losses. hills, Jaintia hills and other smaller hills along the border. The average
Therefore, time and distance are vital elements in emergency response. annual maximum temperature is 23 °C (August to October) and the
2
Fig. 1. Selected study area showing the locations of flood inventory points (main figure). The inset figures on the left show the location of the study area.
average annual minimum temperature is 7 °C (January). Approxi- facility). It is believed that road features considered in this study are a
mately, 80% of the average rainfall (3334 mm/yr) occurs between May highly useful to develop a successful location-allocation model in regard
and September during the monsoon period (Bari et al., 2015). to devising alternative routes as damages of roads after disaster are
common. Likewise, attribute of vehicular speed limits and road type
2.2. Data acquisition will help to manage time for evacuation. These attributes are believed
to be instrumental to advance existing location-allocation modeling,
The use of LAMs for emergency evacuations of flood victims is pro- particularly for data sparse countries. In addition, we established topol-
posed in this study. The Shuttle Radar Topography Mission (SRTM) ogy datasets to eliminate certain topological errors (i.e., overextended
digital elevation model (DEM) data were acquired from https:// line and gap between line) by creating specific topological rules in
earthexplorer.usgs.gov/. Landsat 8 images (Operational Land Imager, ArcGIS (v. 10.5). Finally, a network dataset in a geodatabase was gener-
OLI; and Thermal Infrared Sensor, TIRS) for August 30, 2016 with a res- ated using all the layers. Various connectivity rules among different fea-
olution of 30 m were obtained from the US Geological Survey (https:// tures were determined for ease of access, turns, and attributes of the
earthexplorer.usgs.gov/). MODIS data for March 16, 2015 and August network dataset for optimal calculation. In this study, two types of im-
27, 2015 with resolution of 250 m were acquired from NASA (https:// pedance cutoffs (cost functions) were considered: travel distance
lance-modis.eosdis.nasa.gov/imagery/subsets/) for determining flood (km) and travel time (minutes).
inventory. Rainfall and water level data were collected from the
Bangladesh Water Development Board (BWDB) (https://www. 2.3. Methods
hydrology.bwdb.gov.bd/) between 1991 and 2016. A geological map
(scale 1:100,000) was obtained from the Geological Survey of The seven key steps of the methodological framework of the current
Bangladesh (GSB), while soil type and soil texture maps were obtained study are shown in Fig. 2: (i) flood inventory preparation; (ii) classifica-
(scale 1:100,000) from the Soil Resource Development Institute (SRDI). tion of influencing factors; (iii) feature selection and multi-collinearity
These data are freely available at https://www.barc.gov.bd/. A land test; (iv) modeling process; (v) model performance analysis; (vi)
cover (LC) map with resolution of 300 m was collected from the location-allocation modeling; and (vii) sensitivity analysis of LAMs.
European Space Agency (ESA) (http://maps.elie.ucl.ac.be/CCI/viewer/
download.php). 2.3.1. Preparation of the flood inventory map
To carry out location-allocation analysis, we gathered attribute de- A flood inventory map comprises details of past floods distribution
tails of road features (road type, vehicle speed limits, and barriers), and characteristics (Rahman et al., 2019); this is a prerequisite for
and EECs facilities as point shapes (facility name, coordinates, type of flood susceptibility modeling (Tehrany and Kumar, 2018; Arabameri
3
Fig. 2. Schematic framework of the methodology.
et al., 2019). The flood inventory map was prepared using satellite data Water level data from monitoring stations was used to develop
extracted from Landsat 8 and MODIS, historical flood records, and field flood inundation depth (Fig. S1b) using a method proposed by
surveys. The flood events of 1995, 1998, 2004, 2007, 2014, 2015, and Tingsanchali and Karim (2005). Based on the results of studies
2016 were considered for the development of the flood inventory (Pourghasemi et al., 2019a, 2019b, 2020a, 2020b), and environmental
map. A detailed description of the construction of the flood inventory settings of the study area, a total of 15 flood conditioning factors were
map was given by Rahman et al. (2019). This flood inventory was ran- defined: slope angle, plan curvature (the curvature of the land in a
domly divided into two distinct categories: one for training (70%, 700 horizontal plane), profile curvature (the curvature of the land in a ver-
flood points and 700 non-flood points) and the other for validation pur- tical plane), elevation, normalized difference vegetation index (NDVI),
poses (30%, 300 flood points and 300 non-flood points) (Fig. 1) to aspect, stream power index (SPI), topographic wetness index (TWI),
construct and evaluate the ML and MCDM techniques for flood suscep- drainage density, distance from river, distance from road, soil type,
tibility modeling. soil texture, geology, and LC. Elevation is one of the most important
factors in the assessment of probability of flooding. The chance of
2.3.2. Flood triggering and conditioning factors flooding in low-lying areas is inherently high as water flows from
Rainfall and flood inundation depth are the most important factors high to low-lying areas. This factor was extracted directly from the
affecting whether a flood incident occurs in the current study area. DEM (Fig. 1). Slope is considered another important factor due to its
Thus, these two factors were considered as the triggering factors for direct relation with surface runoff and infiltration. The slope map
flood susceptibility mapping. Rainfall data were derived from eight rain- was extracted from 30-m DEM. Aspect plays an important role in
fall stations in the study area and the Modified Fournier Index (MFI) flood susceptibility mapping because water flow directions can be eas-
(Arnoldus, 1980) was calculated to develop a rainfall thematic map ily calculated from it. The DEM was used to prepare the aspect map.
(Fig. S1a) using the inverse distance weighting (IDW) interpolation Plan and profile curvatures are useful conditioning factors in flood
model (Sarkar and Mondal, 2020). The MFI is computed using Eq. (1) susceptibility mapping from which helpful geomorphologic informa-
(Kourgialas and Karatzas, 2011): tion can be obtained. The TWI is a measure of the concentration of
water at any point. SPI is a measure of the erosive power and intensity
12 R2i of surface runoff. The TWI can be calculated using Eq. (2) and SPI can
MFI ¼ ∑ ð1Þ
i¼1 R be calculated using Eq. (3):

where, Ri is the average monthly rainfall (mm), and R is the average an- AS
TWI ¼ Ln ð2Þ
nual rainfall (mm). tan S
4
SPI ¼ AS tan S ð3Þ 2.3.3. Calculation of number of classes using Shannon's Entropy-
information coefficient
In this study, Shannon's Entropy-information coefficient is proposed
where, AS is the specific catchment area and S is the slope (in degrees) for the calculation of the accurate number of categories of the flood con-
(Khosravi et al., 2019). ditioning factor based on entropy index. The entropy index shows
The drainage density is defined by the proportion of total channel which categories are the most important to the flood susceptibility pre-
lengths to the overall basin area (Siahkamari et al., 2018). The DEM diction. The information coefficient was determined for two to ten cat-
was used to prepare the drainage density map. The distance between egories, for each conditioning factor (i.e., elevation, slope angle, profile
the flood-prone area and the river is another significant factor deter- curvature, SPI, TWI, NDVI, drainage density, distance from river, and dis-
mining occurrence of floods. The distance from river map was prepared tance from road), except for the categorical conditioning factors such as
using the Euclidian Distance method on the basis of the digital network aspect, plan curvature, LC, geological unit, soil type, and soil texture. The
flow layer. The distance from roads were also calculated using the coefficient values vary from 0 to 1. The values close to 0 indicate low ac-
Euclidean Distance method as it is also an important factor in flood sus- curacy and values close to 1 indicate high accuracy. The mathematical
ceptibility mapping. The NDVI was derived from the images collected expressions to calculate the coefficients are presented in Eqs. (5) to
from the Landsat 8 satellite (April 26, 2016) with a resolution of (9) (Constantin et al., 2011):
30 m × 30 m. NDVI is calculated as the ratio between the red (band
Af
4) and near infrared (band 5) values, and for Landsat 8 NDVI was calcu- FPij ¼ ð5Þ
lated using Eq. (4): Ac
FP ij
FPij ¼ ð6Þ
Sj
Band 5−Band 4
NDVI ¼ ð4Þ ∑ FP ij
Band 5 þ Band 4 j¼1
Sj
LC is another significant factor because decreases in vegetation den- Ej ¼ − ∑ FPij log 2 FP ij , j ¼ 1, 2, 3, . . . :, n ð7Þ
sity are correlated with increases in likelihood of flooding (Mojaddadi i¼1
et al., 2017). Bedrock and soil porosity and permeability are also signif-
Ejmax ¼ log 2 Sj ð8Þ
icant factors affecting occurrence of floods. Therefore, bedrock and soil
type are important variables in flood susceptibility modeling (Islam
Ejmax −Ej
and Sado, 2000a; Rahmati et al., 2016). The water infiltration depends Ij ¼ , j ¼ 1, 2, 3, . . . :, n ð9Þ
Ejmax
largely on soil properties, and in particular soil texture. In other
words, the surface runoff and flood inundation are regulated by the
where, Ac is the area percentage of a category, Af is flood percentage of a
soil texture. The list of flood triggering and conditioning factors is
given category, FPij is the probability of flood in the ith class of the jth var-
shown in Table 1.
iable, Ej is the entropy value, Ejmax is the entropy value with S classes,
and Ij is the information coefficient.
Table 1 2.4. Spatial modeling

The flood triggering and conditioning factors.
Factor Data source Reference 2.4.1. Feature selection and multi-collinearity test
Rainfall intensity Rainfall, BWDB Kourgialas and Karatzas (2011), Information Gain (IG) analysis was performed for feature selection.
(MFI)a Sarkar and Mondal (2020) It is one of the most popular methods for determining the most appro-
Depth (m)a Water level, https://www.hydrology.bwdb.gov.bd/ priate flood susceptibility assessment conditioning factors to increase
BWDB
the predictive ability and efficiency of models (Hunt et al., 1966; Javidi
Elevation (m) DEM (30 m) https://earthexplorer.usgs.gov/
Slope (°) DEM (30 m) Samanta et al. (2018), Rahman et al. and Mansoury, 2017).
(2019) This study used multi-collinearity testing (MCT) of the conditioning
Aspect DEM (30 m) Bui et al. (2019) factors to determine whether variables are correlated or not. It can lead
Plan curvature DEM (30 m) Khosravi et al. (2019) to an error in explaining the importance of the model's parameters and
(100/m)
Profile curvature DEM (30 m) Costache et al. (2020)
affect the performance. Generally, a tolerance (TOL) < 0.1 and a vari-
(100/m) ance inflation factor (VIF) > 10 is regarded as a sign of multi-
TWI DEM (30 m) Khosravi et al. (2019), collinearity (Wang et al., 2019).
Pourghasemi et al. (2020b)
SPI DEM (30 m) Khosravi et al. (2019),
2.4.2. Levenberg–Marquardt back propagation neural network
Pourghasemi et al. (2020b)
Drainage density DEM (30 m) Bui et al. (2019) A number of factors should be considered in development of the LM-
(km/km2) BP neural network model; however, there are no rules for selection of
Distance from LGED river https://data.humdata. these factors. MATLAB (v. 2017) with the ‘nftool command’ and
river (km) network org/dataset/bangladesh-water-courses ‘trainlm’ algorithm were used to implement a multi-layer perceptron
Distance from LGED road https://data.humdata.
based LM-BP neural network (Xiong et al., 2019). The most common
road (km) network org/dataset/bangladesh-roads
NDVI Landsat 8 (30 m) https://earthexplorer.usgs.gov/ neural network structure has a total of three possible layers: input, hid-
Land cover (LC) ESA (300 m) https://maps.elie.ucl.ac. den, and output layers. In this work, flood conditioning factors were
be/CCI/viewer/download.php considered as input layers. To predict the susceptibility of flooding and
Geology BARC https://www.barc.gov.bd/
make statistical correlation between the conditioning factors and flood
(scale-1:100,000)
Soil type BARC https://www.barc.gov.bd/ probabilities easier, the data were randomly split into two samples:
(scale-1:100,000) training samples (70%) and testing samples (30%). The training data
Soil texture BARC https://www.barc.gov.bd/ should be different from the validation data in case of developing LM-
(scale-1:100,000) BP. The training data was used to define the key factors in the study
a
Flood triggering factors area which influence flood susceptibility, while testing data was used
5
to check the performance of model. The number of hidden layers ranged

between 1 and 10 and were estimated using trial and error (Elsafi, 1 e 2
σ2 ¼ ∑ T T ð12Þ
2014). e−1 k¼1 jk− j
σ
2.4.3. Decision trees βj ¼ ð13Þ
Tj
The purpose of the DT algorithm is to determine the number of de-
cision rules for predicting the result from a series of input variables. It
is also known as classification and regression trees (CART) when the Substep 3: At this step, a statistical method was implemented to
target variables are discrete or continuous (Debeljak and Džeroski, measure the concordance coefficient (agreement) in the view of experts
2011). The DT models have the ability to predict complex relations which indicate how close the respondent's judgments are from each
among variables (Kheir et al., 2010). In this work, the CART algorithm other. The coefficient of concordance (W) was calculated with Eq. (14):
with ‘RegTree’ command in MATLAB (v. 2015) was employed to build 12S
DT and predict the desired variable (Hateffard et al., 2019). Prior to run- W¼ e
W ∈ ½0; 1 ð14Þ
ning the model, a training (70%) and testing (30%) datasets were gener- e2 ðn3 −nÞ−e ∑ T k
k¼1
ated following procedure mentioned earlier (i.e., LM-BP model).
where, Tk is the index of repeated ranks, e is the total experts and S is the
2.4.4. Extended stepwise weight assessment ratio analysis method total square deviation of the rankings of each criterion. S was calculated
The classical stepwise weight assessment ratio analysis (SWARA) is from Eq. (15):
an MCDM method and was first introduced by Keršuliene et al. (2010).
" #2
Experts play a vital role in the SWARA method in terms of the judgment n e 1 n e
of criteria and weights, and therefore this is called an expert-oriented S ¼ ∑ ∑ T jk− ∑ ∑ T jk ð15Þ
j¼1 k¼1 n j¼1 k¼1
method. The most important criterion according to this method is
ranked 1 and the least important criterion is given the lowest ranking.
Substep 4: This step is used to calculate the importance of the con-
Although it is a comprehensive subjective weighing method, which
cordance coefficient because W is stochastic and therefore an additional
has been extensively used in several fields and has multiple applications 2
statistical parameter (Xα,f) is needed. The significance X2α,f of the concor-
and benefits, this tool requires consistent, precise, and exact decision
dance coefficient at α significance level and the degree of freedom
making (Ghenai et al., 2020). In such situations, a significant difference
f (Zavadskas et al., 2010) was obtained using Eq. (16):
in the weight of the criterion will greatly affect the results, leading to
uncertain and inaccurate ranking or selection. Therefore, in this study, 12S
we have proposed the extended stepwise weight assessment ratio anal- X 2α,f ¼ W:e:ðn−1Þ ¼ e
ð16Þ
ysis (ESWARA) approach for increasing the accuracy of the classical enðn þ 1Þ− n−1
1
∑ Tk
k¼1
SWARA process and minimizing the impact of subjective personal judg-
ments on outcome (Zolfani et al., 2018). In this study, the value of (α) was considered as 0.05. The hypothesis
The following steps define the ESWARA method: of an agreement between independent experts is agreed and group
Step 1: Calculating average attributes value. opinion would be formed, if the value X2α,f is greater than the tabular
The criteria were sorted according to their importance in descending value X2tabl for the degree of importance. Further, if X2α,f < X2tabl then, the
order (Zolfani et al., 2018). The expert assessment method was used agreement of experts is not reasonable and cannot accept the hypothe-
here to determine the importance of each attribute. Then, the average sis of rank correlation.
attributes rank (T j Þ was determined using Eq. (10): Step 3: The factor kj was calculated as follows:
e
∑ T jk kj ¼ 1;j ¼ 1 or kj ¼ sj þ 1;j > 1 ð17Þ
k¼1
Tj ¼ ð10Þ
r
where, Tjk is the rank of the jth criterion by the kth expert and e is the total Step 4: The recalculated weight qj was estimated as follows:
number of experts.
q j−1
Step 2: Identifying sj and the consistency of the expert judgment. qj ¼ 1;j ¼ 1 or qj ¼ ;j > 1 ð18Þ
kj
Step 2 consists of four substeps:
Substep 1: Determining the attribute value in comparison to other
criteria. Keršuliene et al. (2010) referred to this value in the classical
SWARA method as the “comparative importance of average value (sj)” Step 5: The relative weights wj of the evaluation criteria were calcu-
and experts have illustrated the comparative importance of one crite- lated as follows:
rion for the next by setting rough numbers. This value was determined
qj
using the following equation to improve the reliability of the process wk ¼ ð19Þ
n
and to increase the accuracy of the results obtained (Zolfani et al., 2018). ∑ qm
m¼1
Tj
sj ¼ n
ð11Þ n
∑ Tj F P ¼ ∑ wk x ð20Þ
j¼1
j¼1
where, n is the number of factors, wk is the weighting coefficients for

where, Tj is the sum of the attribute's priority values.
conditioning factors, and x is the weighting coefficients for sub-class of
Substep 2: In this step, Eqs. (12), (13) were performed to calculate
each conditioning factor. Finally, the flood probability map was devel-
the experts' ranking dispersion, σ2 and the variation, βj of the results, re-
oped using Eq. (20).
spectively.
6
Initially, all the three models were employed for the calculation of 2003). In terms of allocating the demand points, LSCP model was similar
the flood probability index, which was assigned for the generation of to the MCLP model but varies in relation to the number of facilities allo-
flood probability thematic layers for all pixels in the study area. The cated according to the distance used. It can be derived accordingly
first phase consisted of a unique flood probability index for each pixel (Church and Gerrard, 2003):
in the study area; then, these indices were exported in ArcGIS (v.
10.5) and used to create a final flood probability thematic layer. In the Minimize Z ¼ ∑xj ð23Þ
j∈J
second phase, thematic flood triggering layer was constructed by inte-
grating flood inundation depth with rainfall intensity. Finally, the Subjected to : ∑ pij xj ≥1 for each demand, i ¼ 1, 2, 3 . . . , m
thematic layer of flood probability was converted into a flood suscepti- j∈N i
bility map by multiplying it with the flood triggering layer.

xj ¼ 0, 1 for each facility site, j ¼ 1, 2, 3, . . . , n
2.5. Determination of spatial relationships among conditioning factors and
flood occurrence where, Z is the objective function, pij is the population of demand i, xj is 1
if demand i is allocated to a facility site j and xj is 0 otherwise, Ni repre-
The normalized frequency ratio (nFR) technique was used to deter- sents the collection of facilities when dij ≤ s, dij is the distance between
mine the spatial relationships (Gnyawali et al., 2019). Firstly, we applied i and j, and s is used for important service response time or distance.
the Frequency Ratio (FR) model that can be mathematically expressed
in Eq. (21). Then, the FR value was normalized by the sum and defined 2.6.3. Maximal covering location problem model
as nFR. Larger values of nFR mean that the conditioning factor class is For a certain distance or travel time, the MCLP model seeks to opti-
more significant for flood occurrence. In this way, we determined the mize coverage for the number of possible demand points (Polo et al.,
individual impact of each subclass of the flood determinant. 2015). This model is considered to be suitable when there are a low
number of few facilities and maximum demand must be covered. As
N pix ðF Þ=N
pix ðC Þ
far as facility selection is concerned, the MCLP model does not minimize
FR ¼ ð21Þ the facilities number that are essential to cover all demand points over
∑Npix ðTF Þ=∑N
pix ðTSÞ
certain distances or time, as in the case of the LSCP model (Church
where, Npix(F) are the pixels containing flood in a class, Npix(C) are the and Gerrard, 2003). However, the MCLP model offers solutions that
total pixels of each class of the study area, Npix(TF) are the total pixels cover the largest demand area in compliance with certain criteria for
containing flood, and Npix(TS) are the total pixels of the study area distance, or the time spent between supply and demand. In a number
(Regmi et al., 2014). of studies, the objective function of MCLP has been presented as follows
(Church and ReVelle, 1974):
2.6. Location-allocation models Maximize, Z ¼ ∑pi yi ð24Þ
i∈I
2.6.1. p-Median problem model
The PMP model attempts to minimize the total weighted distance Subjected to : ∑ xi þ yi ≥1 for all i ∈ I
j∈N i
between demand sites and nearest facilities (Polo et al., 2015). In a de-
veloping emergency evacuation, it is vital that the rescuers are able to
∑xj ¼ f
reach the demand point as quickly as possible to prevent loss of j∈J
human life (Romero et al., 2012). The optimization approach of this
model depends on the exchange or replacement of potential locations xj ¼ ð0, 1Þ for all j ∈ J
(chosen and candidate) between demand points (areas of flood-
affected people) and facilities (EECs). The objective function of this yi ¼ ð0, 1Þ for all i ∈ I
model can be specified as follows (Cromley and McLafferty, 2002):
where, Z is the objective function, I is the set of demand nodes, J is the set
Minimize Z ¼ ∑∑pi dij xij ð22Þ of the facility sites, xj = 1 if EEC (facility) is allocated to the site j and
i∈I j∈J
xj = 0 otherwise, pi is the population of demand i, yi = 1 if the flood vic-
tim site (demand) is fulfilled at the site i and yi = 0 otherwise, and f is
Subjected to : xij ≤xjj for all i, j∈J the number of total EECs to be located.
∑xij ≤1 for all i

j∈J
Table 2
∑xjj ¼ f for all j The calculation of information coefficients by Shannon's Entropy index.
j∈J
Factor Information coefficients
xij , xjj ∈f0, 1g Class Class Class Class Class Class Class Class Class
2 3 4 5 6 7 8 9 10
where, Z is the objective function, pi is the population of demand i, dij is Elevation 0.115 0.426 0.148 0.260 0.315 0.279 0.205 0.248 0.209
Slope 0.630 0654 0.677 0.693 0.716 0.739 0.756 0.769 0.779
the travel distance between i and j, xij is 1 if flood victims site (demand Profile 0.038 0.076 0.170 0.187 0.194 0.222 0.215 0.219 0.247
points) i is allocated to a EEC (facility) j and xij is 0 otherwise, and f is the curvature
number of facilities to be used as EECs. TWI 0.017 0.016 0.017 0.020 0.022 0.019 0.021 0.020 0.023
SPI 0.008 0.009 0.010 0.010 0.011 0.010 0.010 0.010 0.016
NDVI 0.030 0.029 0.027 0.028 0.026 0.024 0.023 0.022 0.037
2.6.2. Location set covering problem model
Drainage 0.072 0.080 0.048 0.054 0.032 0.042 0.044 0.034 0.035
The LSCP model intends to minimize distance or time, if the number density
of facilities that cover or satisfy all demand. This model was different Distance 0.004 0.009 0.009 0.009 0.010 0.011 0.012 0.013 0.027
from other models, in that it defines the number of facilities required from river
to meet certain distance requirements to optimally accommodate a cer- Distance 0.008 0.005 0.009 0.010 0.015 0.020 0.026 0.029 0.037
from road
tain number of facilities onto a demanding surface (Church and Gerrard,
7
2.6.4. Maximize attendance model (IG = 0.019), plan curvature (IG = 0.015), and SPI (IG = null). As the
The MA model is designed to optimize the demand attendance with IG value for SPI was null, SPI was excluded from the MLMs for develop-
respect to distance and time spent getting to the point of demand ing flood susceptibility maps.
(Algharib, 2011). This is another form of coverage problem which is dif- The tolerance values (TOL) and variance inflation factor (VIF) were
ferent from previous estimations. The major difference in the MA model calculated by performing multi-collinearity analysis; for all factors
is that it covers a demand weight ratio and does not provide full cover- TOL > 0.1 and VIF < 10 (Table 3). There was no serious multi-
age for each demand level. The findings of this model were helpful for collinearity among the independent variables.
analysis when various demand surfaces were implemented. This
model's objective function is as follows (Holmes et al., 1972): 3.3. Spatial relationship between flood incidence and conditioning factors
n n The spatial relationship was calculated from nFR values of each sub-
Z ¼ ∑ ∑ pi S−dij xij ð25Þ
i¼0 j¼0 class of factors (Table S1). The higher the nFR value, then the higher the
probability of flood occurrence. The nFR value of elevation between 2
where, Z is the objective function, pi is the population of demand i, n is and 12.37 m was 0.687 which was the highest among all other elevation
the flood victim's sites number, S is the threshold distance, dij is the classes. This indicates that the lower the elevation, the higher the flood
travel distance between i and j, and xij = 0 if the flood victim site i is probability is. The estimated value of nFR of slope between 0° and 2°
not covered by a EEC j and 0 < dij ≤ 1 when the victim site i is sheltered was 0.871, which indicates this class is very susceptible to flooding oc-
by a EEC j. currence. On the other hand, nFR value of slope in the class 18° to
31.07° was 0.00, which indicates this class is not susceptible to flood oc-
3. Results and discussion currence. The highest value of nFR (0.161) for aspect was calculated for
the class ‘flat,’ while lowest value of nFR was 0.082 for the class north (0°
3.1. Determining the class numbers of the flood conditioning factors to 22.5°). The highest value of nFR for plan curvature was 0.572 (flat
curvature), which could be explained by the fact that the topography
The maximized information coefficient of each flood conditioning of hillsides direct water flow into lower-lying and flat areas, leading to
factor, represented by the number of classes, was determined by the flooding. The highest nFR value was 0.299 where the profile curvature
Shannon's entropy index according to the procedure discussed in the varied between −0.03 and 0.01 and the lowest nFR value was 0.00
methods. Elevation was categorized by the geometric interval classifica- where profile curvature varied between −1.13 and −0.35. The highest
tion method and the information coefficient was maximized (0.426) nFR (0.148) for the NDVI factor was for the −0.23 to −0.06 class,
when categorized into three classes (Table 2, Fig. S1c). For slope angle, followed by an nFR of 0.142 for the −0.06 to 0.02 class. Increasing vege-
the information coefficient was maximized (0.779) when manually tation density (indicated by an increase in NDVI) was an indicator of de-
classified into ten classes (Fig. S1d). Profile curvature, TWI, SPI, and creasing likelihood of flooding in the study area. The highest nFR was
NDVI were classified into ten classes (Figs. S1h, S1i and S1m) by consid- 0.134, when TWI varied between 24.05 and 28.54, and the lowest nFR
ering the natural break classification method. The maximum informa- was 0.035, where the TWI varied from 0.0 to 2.0. The nFR results on the
tion coefficients for these factors were of 0.247, 0.023, 0.016, and TWI factor showed that in general the nFR value increased as TWI in-
0.037, respectively (Table 2). For drainage density, the information coef- creased, indicating that TWI is directly correlated with the probability
ficient was maximized (0.080) when classified by applying geometric of flood occurrence. The calculation of nFR for the ratio of flood incidence
interval classification into three classes (Fig. S1j). For distance from and drainage density indicated that class 0.72 to 1.40 has the highest nFR
river, the information coefficient was maximized (0.027) when manu- value (0.502), followed by 0.37 to 0.72 (0.330). This indicated that prob-
ally classified into ten classes (Fig. S1k). Finally, for distance from road, ability of flood occurrence increases because of the lower infiltration and
the information coefficient was maximized (0.037) by classifying into higher surface flow with increasing in drainage density. In the case of dis-
ten classes with equal intervals (Table 2, Fig. S1l). tance from river, areas furthest away from the river (> 18 km) showed
the lowest probability of flood occurrence. In general, probability of
3.2. Feature selection and multi-collinearity assessment of flood condition- flooding increased with decreasing distance from the river. Distance
ing factors from the road is another important factor affecting the likelihood of
flooding. Impervious surfaces such as asphalt and the surrounding
IG analysis, to assess the significance of each conditioning factor in areas have a significant influence on flood occurrence because cannot
susceptibility assessment (Table 3), showed that slope angle (IG = penetrate the surface. The distance from road in the class 15.79 to
0.164) had the highest impact on flooding. This was followed by geology 18.04 km had the highest nFR value of 0.125, indicating a higher probabil-
(IG = 0.140), soil texture (IG = 0.133), soil type (IG = 0.121), elevation ity of flooding; in contrast, the distance from road varied between 20.30
(IG = 0.104), LC (IG = 0.081), distance from river (IG = 0.070), TWI and 22.56 km had the lowest nFR value of 0.018, indicating that areas far
(IG = 0.063), drainage density (IG = 0.060), profile curvature (IG = from roads were least prone to flooding. With the exception of water
0.049), aspect (IG = 0.048), NDVI (IG = 0.026), distance from road bodies, the highest nFR value for LC classes was for grassland (0.243),
Table 3
Multi-collinearity analysis of flood conditioning factors.
Sl. no. Conditioning factors IG Collinearity statistics Sl. no. Conditioning factors IG Collinearity statistics
Tolerance VIF Tolerance VIF
1 Elevation 0.104 0.540 1.851 9 Drainage density 0.060 0.742 1.347

2 Slope 0.164 0.734 1.363 10 Distance from river 0.070 0.876 1.142
3 Aspect 0.048 0.838 1.193 11 Distance from road 0.019 0.954 1.049
4 Plan curvature 0.015 0.582 1.719 12 Land cover (LC) 0.081 0.558 1.793
5 Profile curvature 0.049 0.578 1.730 13 Soil type 0.121 0.409 2.442
6 NDVI 0.026 0.809 1.236 14 Soil texture 0.133 0.832 1.201
7 SPI Null 0.677 1.478 15 Geology 0.140 0.367 2.727
8 TWI 0.063 0.592 1.688 – – – –
8
3.4. Preparation of flood susceptibility maps
In the training of LM-BP model, the neural network was stopped

after 120 iterations. The precision objective was reached by an RMSE
value of 8.532 and R2 value of 0.785 from the neural network structure
of 15 input layers–10 hidden layers–1 output layer. This corresponds to
our expectations regarding the accuracy of the flood probability
assessment.
In the DT model, logical guidelines were used to predict the proba-
bilities of the flood via a tree-like structure. The RMSE was 7.942 and
R2 was 0.755. This reflects our expectations about the accuracy of the as-
sessment of flood probability.
In the ESWARA model, the weight coefficients were calculated to as-
sess the flood probability. The higher the weights, the greater the effect
of a particular factor. Of the different conditioning factors, slope
(0.1091), elevation (0.0977), geology (0.0879), profile curvature
(0.0796), and aspect (0.0731) had the most significant influence on
the probability of flood incidence in the study area (Fig. 3). The condi-
Fig. 3. Weight of flood conditioning factors using ESWARA model.
tioning factors SPI (0.0676), LC (0.0632), NDVI (0.0591), drainage den-
sity (0.0575), TWI (0.0548), soil texture (0.0523), and distance from
indicating a very high risk of flooding. For soil types, the highest nFR river (0.0509) were intermediate (Fig. 3). The conditioning factors soil
values were recorded non-calcareous alluvium (0.212) and peat type (0.0496), distance from road (0.0490), and plan curvature
(0.208), and the lowest value was recorded for the deep red brown ter- (0.0485) had the least influence on the probability of flood occurrence
race soil (0.004). As for soil texture, the highest nFR value was recorded (Fig. 3). The maps were grouped into four categories: low susceptibility,
for mixed silt loam and silty clay loam soil (2.03), and the lowest value medium susceptibility, high susceptibility, very high susceptibility con-
was recorded for mixed gravelly clay loam and silty clay (0.007). Analysis sidering the natural breaks classification method (Achour and
of the nFR for the spatial relationship between flood probability and geo- Pourghasemi, 2019; Mirchooli et al., 2019; Mohammady et al., 2019;
logical units, show that (with the exception of water bodies), the highest Pourghasemi et al., 2020) shown in Fig. 4 and the areas coverage for
nFR value was for marsh clay and peat (0.208). each class is presented in Fig. 5.
Fig. 4. Flood susceptibility maps: (a) LM-BP, (b) ESWARA, (c) DT, and (d) population risk map.
9
and map accuracy quantitatively. Yesilnacar and Topal (2005) men-

tioned that the AUROC and prediction accuracy relationship varies
from 0 to 1 which was classified into five categories: excellent, >0.9 to
1; very good, >0.8 to 0.9; good, >0.7 to 0.8; fair, >0.6 to 0.7; and poor,
0.5 to 0.6. For the training data, the values of the AUROC curve for the
LM-BP, DT, and ESWARA models were 0.904, 0.884, and 0.859, respec-
tively (Fig. 6a). For the test data, the values of the AUROC curve for the
LM-BP, DT, and ESWARA models were 0.856, 0.826, and 0.809, respec-
tively (Fig. 6b). Therefore, it can be said that the LM-BP had a greater
predictive capability than the DT and ESWARA models. These findings
correspond to recent studies (Kia et al., 2012; Falah et al., 2019). The
findings of this research can be useful for choosing the most effective
flood spatial modeling in the study area and elsewhere.
3.6. Preparation of location-allocation models
The first step toward implementing the LAM is to locate existing

EECs and demand points (flood victims). Thus, considering the flood
susceptibility map and population density map, a flood victim's risk
map was prepared to locate where the EECs for flood victims are re-
Fig. 5. Flood susceptibility percentage of each class for the developed maps. quired. Bangladesh is one of the world's most populated countries and
the majority of the population is located in the major cities. Digital pop-
ulation data were generated using the Sylhet divisional population
In comparison with DT and ESWARA, the LM-BP model produced census 2011 (https://www.bbs.gov.bd/). Based on the density of popu-
better performance. However, some higher elevation areas were classi- lation, the digital population data were categorized into four classes:
fied by the LM-BP method as having a medium or high risk of flooding, (i) 1–500 per km2 (low); (ii) 501–700 per km2 (medium); (iii)
whereas the DT and ESWARA models identified these areas as having a 701–900 per km2 (high); and (iv) >900 per km2 (very high). Finally,
lower risk of flooding. In general, the results from three models were in the most accurate flood map (LM-BP) was incorporated with the digital
agreement concerning which areas had a low probability of flooding population density map and then, ranking matrix of the two-
and which areas had a very high probability of flooding. Further the dimensional multiplication mode was applied to develop a flood
comparison results showed that, three models have the similar predic- victim's risk map (Fig. 4d). The ranks of the risk map ranged from 1 to
tion results of 56.82%. Since no decisions can be made without the eval- 16. The higher the rank, the higher the risk. For clarity, the ranks were
uation statistics, validation was performed by plotting the AUROC curve, reclassified into four categories: low (1–3); medium (4–6); high
described in the following section. (7–11); very high (12–16) (Frigerio et al., 2016). Some areas that had
a high susceptibility to flooding were assessed as low risk because of
3.5. Model validation the low population density, whereas some areas that had a high suscep-
tibility to flooding were assessed as high risk because of the high popu-
In this study, the AUROC curve was used to plot and evaluate the lation density. Finally, for location-allocation modeling purposes,
prediction ability of standalone models. The AUROC is a graphical repre- the demand points were regarded as the centroid points of the
sentation created by plotting the sensitivity on the x-axis against the 1- 500 m × 500 m grid cell in the risk map. A total of 62,147 centroid
specificity on the y-axis. The AUROC shows model prediction capability grid points was listed as demand points and 99 existing EECs have
Fig. 6. Model validation of flood susceptibility: (a) training data, (b) testing data.
10
Table 4
Calculated demand points for EECs using the p-median problem (PMP) and MA models (Hospital*, Fire station**, College***, and School****).
EECs No. Demand Demand Impedance cutoff Total weight EECs No. Demand Demand Impedance cutoff Total weight
points weight (minutes) points weight (minutes)
PMP MA PMP MA PMP MA PMP MA
EEC 01* 3 3 2.99 0.77 0.77 0.77 EEC 51**** 417 417 304.31 6761.52 6761.52 4271.07
EEC 02**** 8 8 7.53 28.18 28.18 26.15 EEC 52**** 589 589 502.29 5202.65 5202.65 4279.99
EEC 03**** 257 257 223.53 2007.96 2007.96 1642.31 EEC 53**** 2140 2140 1333.18 48,409.33 48,409.33 24,894.74
EEC 04**** 271 271 232.12 2333.08 2333.08 1903.95 EEC 54**** 1496 1496 1052.29 26,622.71 26,622.71 15,285.84
EEC 05* 632 632 506.90 7505.81 7505.81 5729.39 EEC 55**** 858 858 645.96 12,722.21 12,722.21 8310.69
EEC 06* 1366 1366 811.39 33,276.41 33,276.41 13,042.39 EEC 56**** 2019 2019 1206.26 48,764.56 48,764.56 21,727.20
EEC 07* 1381 1381 993.34 23,259.81 23,259.81 14,886.00 EEC 57** 151 151 129.81 1271.14 1271.14 1016.88
EEC 08* 220 220 173.36 2798.61 2798.61 2043.12 EEC 58**** 976 976 685.71 17,417.42 17,417.42 11,143.85
EEC 09* 1669 1669 1144.59 31,464.46 31,464.46 17,831.50 EEC 59**** 45 45 41.62 202.54 202.54 182.07
EEC 10* 2033 2033 1418.13 36,891.91 36,891.91 23,247.77 EEC 60*** 1104 1104 831.69 16,338.87 16,338.87 10,493.51
EEC 11* 1171 1171 942.25 13,725.17 13,725.17 10,445.77 EEC 61*** 711 711 549.11 9713.57 9713.57 7023.09
EEC 12* 6 6 5.92 4.85 4.85 4.78 EEC 62**** 413 413 323.36 5378.23 5378.23 3777.76
EEC 13* 1068 1068 864.37 12,218.08 12,218.08 9354.80 EEC 63**** 2 2 1.99 0.61 0.61 0.61
EEC 14* 1475 1475 1149.63 19,521.98 19,521.98 14,341.46 EEC 64**** 461 461 381.04 4797.43 4797.43 3804.55
EEC 15* 605 605 529.50 4530.23 4530.23 3726.87 EEC 65**** 88 88 79.24 525.81 525.81 458.29
EEC 16* 841 841 594.24 14,805.36 14,805.36 8983.74 EEC 66**** 3 3 2.99 0.30 0.30 0.30
EEC 17* 2776 2776 1767.63 60,502.42 60,502.42 30,849.51 EEC 67**** 48 48 45.28 163.33 163.33 150.99
EEC 18* 1056 1056 768.94 17,223.44 17,223.44 11,698.57 EEC 68**** 11 11 10.59 24.51 24.51 23.26
EEC 19* 456 456 371.71 5057.59 5057.59 3852.25 EEC 69**** 512 512 411.11 6053.25 6053.25 4538.17
EEC 20* 1278 1278 770.42 30,454.54 30,454.54 12,934.08 EEC 70**** 492 492 362.43 7774.41 7774.41 4995.24
EEC 21* 3457 3457 1973.58 89,005.03 89,005.03 42,036.62 EEC 71**** 249 249 202.77 2773.75 2773.75 2068.66
EEC 22* 2166 2166 1501.92 39,844.97 39,844.97 24,896.50 EEC 72*** 25 25 24.60 24.17 24.17 23.63
EEC 23* 2302 2302 1804.62 29,842.87 29,842.87 22,061.92 EEC 73**** 406 406 196.99 12,540.70 12,540.70 4039.62
EEC 24* 1 1 1.00 0.09 0.09 0.09 EEC 74**** 248 248 206.31 2501.31 2501.31 1966.46
EEC 25* 1119 1119 931.78 11,233.47 11,233.47 8994.36 EEC 75**** 40 40 37.67 139.65 139.65 129.91
EEC 26* 854 854 623.27 13,844.08 13,844.08 8939.96 EEC 76**** 0 0 0.00 0.00 0.00 0.00
EEC 27* 59 59 49.74 555.59 555.59 447.91 EEC 77**** 22 22 21.66 20.39 20.39 19.77
EEC 28* 3575 3575 2030.88 92,647.36 92,647.36 36,638.93 EEC 78**** 270 270 214.17 3349.65 3349.65 2503.92
EEC 29**** 4 4 3.98 1.19 1.19 1.19 EEC 79**** 192 192 169.16 1370.62 1370.62 1143.09
EEC 30**** 1 1 1.00 0.07 0.07 0.07 EEC 80**** 14 14 13.72 16.54 16.54 16.15
EEC 31**** 3 3 2.98 1.04 1.04 1.03 EEC 81*** 0 0 0.00 0.00 0.00 0.00
EEC 32* 2 2 1.99 0.49 0.49 0.49 EEC 82**** 360 360 297.37 3757.88 3757.88 2918.15
EEC 33*** 69 69 61.71 437.56 437.56 383.69 EEC 83**** 7 7 6.88 6.91 6.91 6.76
EEC 34**** 1 1 1.00 0.24 0.24 0.24 EEC 84**** 2 2 1.99 0.89 0.89 0.88
EEC 35* 0 0 0.00 0.00 0.00 0.00 EEC 85* 0 0 0.00 0.00 0.00 0.00
EEC 36*** 0 0 0.00 0.00 0.00 0.00 EEC 86**** 1 1 1.00 0.30 0.30 0.29
EEC 37**** 0 0 0.00 0.00 0.00 0.00 EEC 87*** 419 419 326.55 5546.99 5546.99 3724.53
EEC 38* 1012 1012 656.52 21,328.51 21,328.51 10,928.45 EEC 88* 1378 1378 557.95 49,203.18 49,203.18 14,202.43
EEC 39* 5 5 4.97 1.67 1.67 1.65 EEC 89* 148 148 138.17 589.72 589.72 534.30
EEC 40** 575 575 453.13 7312.47 7312.47 5351.59 EEC 90**** 136 136 111.78 1453.10 1453.10 1082.55
EEC 41** 1838 1838 1317.32 31,241.08 31,241.08 18,035.97 EEC 91* 423 423 336.71 5177.46 5177.46 4012.93
EEC 42**** 520 520 446.86 4388.59 4388.59 3627.35 EEC 92* 1124 1124 735.75 23,295.04 23,295.04 13,220.46
EEC 43**** 564 564 481.02 4979.05 4979.05 3994.89 EEC 93* 16 16 15.47 31.70 31.70 30.44
EEC 44**** 252 252 202.37 2978.09 2978.09 2179.58 EEC 94**** 0 0 0.00 0.00 0.00 0.00
EEC 45* 604 604 491.48 6751.38 6751.38 5160.54 EEC 95* 18 18 17.71 17.32 17.32 16.95
EEC 46**** 389 389 291.88 5827.35 5827.35 4059.47 EEC 96* 0 0 0.00 0.00 0.00 0.00
EEC 47**** 906 906 767.82 8290.95 8290.95 6549.65 EEC 97**** 118 118 104.31 821.30 821.30 701.15
EEC 48*** 142 142 128.95 783.10 783.10 689.02 EEC 98**** 2 2 1.99 0.65 0.65 0.64
EEC 49** 672 672 560.68 6679.37 6679.37 5249.31 EEC 99**** 388 388 341.91 2765.59 2765.59 2384.85
EEC 50*** 44 44 40.78 193.36 193.36 177.14 – – – – – – –
been selected as facility points for location-allocation modeling. The were served by the existing EECs where EEC 35 received the largest
reason for using 500 m × 500 m grid size for population is because of number of demand points (3575) (Table 4).
its availability. Moreover, capacity for EECs are not considered in this The MCLP model showed similar results in terms of demand selec-
paper due to its focus on reducing travel time to evacuate the most vul- tion (Alshwesh, 2014). The correlation between the results of the PMP
nerable people therefore, larger grid size assumes to have less influence and MCLP models was the same because of their similar features in min-
in our work. imizing the overall weighted distance between the EECs and flood vic-
In the second step, a detailed and accurate map of the road network tim sites (demand points), and with regard to the handling of
was constructed which is consistent with the actual situation of roads demand. Maximum distance constraints were introduced for the PMP
including junctions, lengths and connections to implement in the model. Our results are in line with a previous study (Algharib, 2011).
LAM. Then, the PMP model was parameterized to select the optimum However, it is noted that a total of eight existing EECs (i.e. EECs 35, 36,
EECs to serve the maximum demand points within the specified cover- 37, 76, 81, 85, 94, and 96) did not serve any demand points. At the
age time of 60 min from the closest facilities. In this study, the weights of same time, several EECs were located far away from the demand and
all EECs were decided according to their importance. We have consid- had little coverage of the demand points (Table 4 and Fig. 7c).
ered that all the EECs have the same importance so the demand weight The LSCP model was applied to minimize the distances and
would be same for all EECs, i.e., 1 in this case. The results of the PMP determine the minimum number of EECs required for serving all
model showed that a total of 58,250 demand points out of 62,147 demand points within the defined coverage distance or travel
11
Fig. 7. (a) Road network and existing EECs, (b) demand points (low to very high), and (c) minimize weighted impedance or p-median problem (PMP) model.
time. Table 5 shows the calculated demand points for EECs using Finally, as the study hypothesized that the EECs need to be located so
the LSCP model. In total, 15 EECs serve 58,250 demand points; that all the flood victim sites are covered within the specified coverage
EEC 38 served the most demand points (9006), while EEC 19 time of 60 min from the closest facility and maximum demand points
served the fewest demand points (456). It is noted that, most would be served within a 10-min drive of existing EECs. Thus, this
EECs were overloaded; therefore, it could be difficult or costly study considers different impedance factors using the MCLP model to as-
for authorities to serve flood-affected people during floods be- sess the demand point coverage within the impedance cutoffs (Fig. 8). In
cause of the uneven and inadequate distribution of EECs. this respect, a sensitivity analysis was performed based on several im-
The MA model was applied to maximize the attendance of demand pedance thresholds to check whether the impedance could meet the re-
for each EEC and to minimize the total distance between the EECs and quirements for the selection of current EECs to reach all demand points
the flood victim demand points. In terms of allocation of EECs and (Table 6). The findings show that 71.89% of the demand points were inac-
choosing the demand, the MA model is different from the PMP, MCLP, cessible within the specified coverage time of 10 min from the EECs
and LSCP models. The results showed that the total demand weight (Table 6). The percentage of inaccessible demand points decreased
for the MA model was less than that for the PMP model when providing with the increased of impedance cutoffs. The target demand points
99 facilities (EECs) within the distance specified (Table 4). The reason is (62,147 points) coverage within the targeted critical time of 60 min
that the MA model uses the ratio of the total demand weight per de- was 93.73% (58,250 points) which did not meet the objective of the
mand point; the ratio decreased when the distance between the EECs study. Therefore, it can be concluded that the study area required more
and the demand points increased. EECs to cover the entire demand points within the critical time of 60 min.
Table 5
Calculated demand points for EECs using the LSCP model (Hospital*, Fire station**, and School****).
EECs Demand Demand Impedance cutoff Total EECs No. Demand Demand Impedance cutoff Total weight
no. points weight (minutes) weight points weight (minutes)
EEC 06* 1371 1371 33,461.943 33,461.943 EEC 38* 9006 9006 209,607.72 209,607.72
EEC 07* 1828 1828 35,219.839 35,219.839 EEC 41** 7727 7727 197,152.85 197,152.85
EEC 10* 4594 4594 127,937.39 127,937.39 EEC 45* 3003 3003 64,141.317 64,141.317
EEC 17* 3154 3154 71,359.071 71,359.071 EEC 53**** 4995 4995 114,583.23 114,583.23
EEC 19* 456 456 5057.5859 5057.5859 EEC 55**** 5002 5002 137,043.96 137,043.96
EEC 20* 2127 2127 63,519.948 63,519.948 EEC 73**** 2767 2767 87,206.386 87,206.386
EEC 23* 5232 5232 101,619.14 101,619.14 EEC 88* 1662 1662 56,104.962 56,104.962
EEC 28* 5326 5326 142,805.61 142,805.61 – – – – –
12
Fig. 8. Results of applying the MCLP model considering different impedance cutoffs: (a) 10 min. Impedance; (b) 20 min. Impedance; (c) 30 min. Impedance; (d) 40 min. Impedance;
(e) 50 min. Impedance; and (f) 60 min. Impedance.
Table 6
Results of the sensitivity analysis of the MCLP model considering different impedance cutoffs.
Population risk category due to flood Demand points (Nos.) Percentage (%) Impedance cutoffs (in minutes) and unreachable demand points of EECs
10 20 30 40 50 60
Low 28,172 45.33 −20,509 −9973 −5496 −2998 −1633 −1077

Medium 26,765 43.07 −18,780 −10,312 −6136 −4457 −3232 −2507
High 4890 7.87 −3581 −1817 −732 −448 −313 −191
Very high 2320 3.73 −1810 −1012 −638 −468 −352 −122
Total 62,147 100 −44,680 −23,114 −13,002 −8371 −5530 −3897
– – Percentage (%) 71.89 37.19 20.92 13.47 8.90 6.27
13
existing EECs. Therefore, more EECs would be required to provide ac-

ceptable service to the flood victims. From the MCLP, MA, and PMP
models, it was found that there were eight EECs that could not be
reached by flood victims within the specified travel time. Keeping in
mind the objective of the study, an optimized MCLP model was applied
to select the optimal distribution of EECs across the study area and re-
distribute the existing EECs to ensure a better coverage of demand
points. In this study, the optimization of the MCLP model was per-
formed using the PMP model as the optimization of facilities using the
single MCLP approach could be difficult because of the uncertainty
(Romero et al., 2012). In the optimization process, we identified 21 tem-
porary EECs that are needed to serve the demand and fulfill the objec-
tive of the study. In addition, eight existing EECs mentioned in the
earlier section were relocated by the PMP model because the demand
points were not sufficiently covered and located to the areas which
were beyond the standard impedance cutoff. Moreover, the binary
method was replaced with partial coverage in the optimized MCLP
model because the dynamic or partial coverage models had substan-
tially better coverage for demand points in LAMs than the binary
method alone (Reduction, 2004; Tong and Murray, 2009; Lamovec
Fig. 9. Comparison of the results of the MCLP model for the existing EECs against new EECs et al., 2013).
considering different impedance cutoffs. To reduce the overall demand weighted distance between EECs and
demand points and demand coverage, the proposed spatial distribution
of EECs in the study was much better than the current EECs locations.
3.7. Application of optimized maximal covering location problem model to This represents advantages of using LAMs for the development of new
select the best distribution for facility locations facilities such as EECs. Furthermore, the results showed that 47.88% of
demands points could be reached within a travel distance of 10 min to
After comparing the four LAMs, it was observed that, the existing 99 serve the flood-affected people while the previous analysis showed a
EECs were insufficient to access all the demand points within the spec- coverage of 28.18%. This percentage increases when the impedance cut-
ified travel time of 60 min. The results indicated that 58,250 of 62,147 off is increased for redistribution of EECs along with new EECs. The
demand points (population) (93.73%) are within a 60-min drive of 60 min impedance cutoff showed the maximum demand points
Fig. 10. The results of the optimized MCLP model after redistributing and including new EECs.
14
(62,147 out of 62,147 demand points) (Figs. 9 and 10b). This result is Acknowledgements
feasible to fulfill the objective of our study which can further be opti-
mized by considering capacity of each EEC. If the capacity of each EEC This study is based on a PhD thesis by first author: Mahfuzur
is known then integrated model could easily determine whether new Rahman, which is supervised by Professor Ningsheng Chen. The authors
EECs are required or not. acknowledge and appreciate the provision of rainfall and water level
Furthermore, due to an increase in the need of location-allocation data by the Bangladesh Water Development Board (BWDB), without
modeling for emergency evacuation planning in the frequently flooded which this study would not have been possible. Authors would like to
regions such as Bangladesh, we believe that this research could add sig- thank the handling editor for his editorial comments and anonymous
nificant value to the field. Especially, when there has been very limited reviewers for their careful review of the original manuscript and their
research regarding location-allocation modeling for emergency evacua- valuable suggestions which helped us to improve the manuscript. This
tions, researchers could use the proposed work as a foundation for fur- research was funded by the National Natural Science Foundation of
ther work in the field. Finally, it can be concluded that these results help China (Grant Nos. 41861134008 and 41671112) and the 135 Strategic
to fill noticeable service gaps of the study area. Program of the Institute of Mountain Hazards and Environment (IMHE),
Chinese Academy of Sciences (CAS) (Grant No. SDS-135-1705).
4. Conclusion
Appendix A. Supplementary data
In this study, we have proposed flood susceptibility modeling inte-
grated with LAMs to facilitate EECs for flood-affected people. Firstly, Supplementary data to this article can be found online at https://doi.
flood susceptibility assessments were carried out using LM-BP, DT, org/10.1016/j.gsf.2020.09.022.
and ESWARA models. Of the three standalone models, LM-BP had the
highest AUROC values (0.904 for the training data, and 0.856 for the References
test data). Therefore, the resulting map produced from LM-BP model
Achour, Y., Pourghasemi, H.R., 2019. How do machine learning techniques help in increas-
was considered as the most accurate flood map for the study area. Sec- ing accuracy of landslide susceptibility maps? Geosci. Front. 11 (3), 871–883.
ondly, a population risk map was developed from the LM-BP flood map Alam, M., Ali, M.H., 2002. Concept of Flood Shelter to Cope With Flood: Engineering Con-
and digital population data using a two-dimensional matrix multiplica- cerns of Flood. Bangladesh University of Engineering and Technology (BUET), Dhaka,
pp. 175–186.
tion method.
Algharib, S.M., 2011. Distance and Coverage: An Assessment of Location-allocation
Finally, four LAMs—PMP, MCLP, LSCP, and MA—were proposed to Models for Fire Stations in Kuwait City, Kuwait. Ph.D. thesis. Kent State University.
serve the demand points (flood victims i.e., children, disabled and el- Ali, S.A., Khatun, R., Ahmad, A., Ahmad, S.N., 2019. Application of GIS-based analytic hier-
derly) through the existing EECs within the specified travel time of archy process and frequency ratio model to flood vulnerable mapping and risk area
estimation at Sundarban region, India. Model. Earth Syst. Environ. 5 (3), 1083–1102.
60 min. The demand points for LAMs were extracted from the devel- Al-Juaidi, A.E., Nassar, A.M., Al-Juaidi, O.E., 2018. Evaluation of flood susceptibility map-
oped population risk map (low to very high). The findings of the pro- ping using logistic regression and GIS conditioning factors. Arab. J. Geosci. 11 (24),
posed models indicate that 6.27% of the demand points are 765.
Alshwesh, I.O.A., 2014. GIS-based Interaction of Location Allocation Models With Areal In-
inaccessible within the specified time of 60 min from the existing
terpolation Techniques. Ph.D. thesis. University of Leicester.
EECs. The sensitivity analysis of the MCLP model indicated that very Anees, M.T., Abdullah, K., Nawawi, M., Ab Rahman, N.N.N., Piah, A.R.M., Zakaria, N.A.,
high-risk areas need to pay more attention in terms of rescue manage- Syakir, M., Omar, A.M., 2016. Numerical modeling techniques for flood analysis.
ment. Therefore, in this study MCLP model was optimized using the J. Afr. Earth Sci. 124, 478–486.
Arabameri, A., Rezaei, K., Cerdà, A., Conoscenti, C., Kalantari, Z., 2019. A comparison of sta-
PMP model for considering partial area coverage instead of binary anal- tistical methods and multi-criteria decision making to map flood hazard susceptibil-
ysis. We relocated eight existing EECs and proposed 21 new temporary ity in Northern Iran. Sci. Total Environ. 660, 443–458.
EECs (to be located during a flood) into their new optimal positions con- Armah, F.A., Yawson, D.O., Yengoh, G.T., Odoi, J.O., Afrifa, E.K., 2010. Impact of floods on
livelihoods and vulnerability of natural resource dependent communities in Northern
sidering the optimized MCLP model. It was observed that new facilities Ghana. Water 2 (2), 120–139.
(EECs) receive all the demand points within the specified time of Arnoldus, H., 1980. An approximation of the rainfall factor in the. Univers. Soil Loss Eq.
60 min. Furthermore, the demand point coverage increased to 29,757 127-132.
(47.88%) from 17,511 (28.18%) when a 10 min impedance cutoff was Bari, S., Rahman, M., Hussain, M., Ray, S., 2015. Forecasting monthly precipitation in Sylhet
city using ARIMA model. Civil Environ. Res. 7 (1), 69–77.
initiated for EECs. Therefore, it can be concluded that around 48% BBS, 2019. Gender Statistics of Bangladesh, 2018. Bangladesh Bureau of Statistics (BBS).
flood victims would be evacuated from the nearest EECs in 10 min Bermúdez, M., Cea, L., Puertas, J., 2019. A rapid flood inundation model for hazard map-
after occurrence of a flood. ping based on least squares support vector machine regression. J. Flood Risk
Manag. 12, e12522.
It can be concluded that, the key innovation of the present study is
Bui, D.T., Ngo, P.T.T., Pham, T.D., Jaafari, A., Minh, N.Q., Hoa, P.V., Samui, P., 2019. A novel
the use of the four LAMs to serve the flood victims. The resulting map hybrid approach based on a swarm intelligence optimized extreme learning machine
could provide measures to reduce susceptibility to flooding and assist for flash flood susceptibility mapping. Catena 179, 184–196.
the affected population prior to floods. Furthermore, the findings of Cabrera, J.S., Lee, H.S., 2019. Flood-prone area assessment using GIS-based multi-criteria
analysis: a case study in Davao Oriental, Philippines. Water 11 (11), 2203.
this study can be beneficial for policymakers, flood mitigation depart- CEGIS, 2012. Master plan of the Haor area, Bangladesh Haor and Wetland Development
ment and emergency evacuation organizations for planning the spatial Board, Ministry of Water Resources. The People’s Republic of Bangladesh 2.
distribution of EECs for flood victims. In addition, this work addressed Chang, M.S., Tseng, Y.L., Chen, J.W., 2007. A scenario planning approach for the flood
emergency logistics preparation problem under uncertainty. Transp. Res. E 43 (6),
a first-hand work for both pre-and post-disaster management planning, 737–754.
and it is not limited to flood disaster. The method can also be applied to Chang, M.J., Chang, H.K., Chen, Y.C., Lin, G.F., Chen, P.A., Lai, J.S., Tan, Y.C., 2018. A support
other natural disasters i.e., landslide, earthquake, and cyclone, etc. In fu- vector machine forecasting model for typhoon flood inundation mapping and early
flood warning systems. Water 10 (12), 1734.
ture studies, we will try to incorporate the LAMs with hydrodynamic
Chapi, K., Singh, V.P., Shirzadi, A., Shahabi, H., Bui, D.T., Pham, B.T., Khosravi, K., 2017. A
models, i.e., HEC RAS, FLO-2D. Future studies could utilize multi- novel hybrid artificial intelligence approach for flood susceptibility assessment. Envi-
objective location-allocation problem to consider other LAMs and ca- ron. Model Softw. 95, 229–245.
pacity of individual EEC to refine finding of this work. Chowdhmy, J.U., Karim, M.P., 1996. A risk-based zoning of storm surge prone area of the
Ganges Tidal Plain. J. Civil Eng. 24 (2), 221–233.
Church, R.L., Gerrard, R.A., 2003. The multi-level location set covering model. Geogr. Anal.
Declaration of Competing Interest 35 (4), 277–289.
Church, R., ReVelle, C., 1974. The maximal covering location problem. Papers of the Re-
The authors declare that they have no known competing financial gional Science Association, Springer-Verlag 32 (1), 101–118.
Constantin, M., Bednarik, M., Jurchescu, M.C., Vlaicu, M., 2011. Landslide susceptibility as-
interests or personal relationships that could have appeared to influ- sessment using the bivariate statistical analysis and the index of entropy in the
ence the work reported in this paper. Sibiciu Basin (Romania). Environ. Earth Sci. 63 (2), 397–406.
15
Costache, R., Hong, H., Pham, Q.B., 2020. Comparative assessment of the flash-flood po- and weights-of-evidence bivariate statistical models with multi-criteria decision-
tential within small mountain catchments using bivariate statistics and their novel making technique. Nat. Hazards 83 (2), 947–987.
hybrid integration with machine learning models. Sci. Total Environ. 711, 134514. Khosravi, K., Shahabi, H., Pham, B.T., Adamowski, J., Shirzadi, A., Pradhan, B., Dou, J., Ly, H.-
Cromley, E., McLafferty, S., 2002. Analyzing access to health services. GIS Public Health 2, B., Gróf, G., Ho, H.L., 2019. A comparative assessment of flood susceptibility modeling
303–337. using multi-criteria decision-making analysis and machine learning methods.
Debeljak, M., Džeroski, S., 2011. Decision trees in ecological modelling. Modelling com- J. Hydrol. 573, 311–323.
plex Ecological Dynamics. Springer 197-209. Kia, M.B., Pirasteh, S., Pradhan, B., Mahmud, A.R., Sulaiman, W.N.A., Moradi, A., 2012. An
Dewan, A.M., Islam, M.M., Kumamoto, T., Nishigaki, M., 2007. Evaluating flood hazard for artificial neural network model for flood simulation using GIS: Johor River Basin,
land-use planning in Greater Dhaka of Bangladesh using remote sensing and GIS Malaysia. Environ. Earth Sci. 67 (1), 251–264.
techniques. Water Resour. Manag. 21 (9), 1601. Kim, T.H., Kim, B., Han, K.Y., 2019. Application of fuzzy TOPSIS to flood hazard mapping for
Dinu, S., Ciucur, V., 2016. Location-allocation models and new solution methodologies in levee failure. Water 11 (3), 592.
telecommunication networks. MS&E 145 (8), 082021. Kourgialas, N.N., Karatzas, G.P., 2011. Flood management and a GIS modelling method to
Elsafi, S.H., 2014. Artificial neural networks (ANNs) for flood forecasting at Dongola Sta- assess flood-hazard areas—a case study. Hydrol. Sci. J. 56 (2), 212–225.
tion in the River Nile, Sudan. Alexandr. Eng. J. 53 (3), 655–662. Lamovec, P., Velkanovski, T., Mikos, M., Osir, K., 2013. Detecting flooded areas with ma-
Erena, S.H., Worku, H., De Paola, F., 2018. Flood hazard mapping using FLO-2D and local chine learning techniques: case study of the Selška Sora river flash flood in Septem-
management strategies of Dire Dawa city, Ethiopia. J. Hydrol. 19, 224–239. ber 2007. J. Appl. Remote Sens. 7 (1), 073564.
Falah, F., Rahmati, O., Rostami, M., Ahmadisharaf, E., Daliakopoulos, I.N., Pourghasemi, Lee, S., Kim, J.C., Jung, H.S., Lee, M.J., Lee, S., 2017. Spatial prediction of flood susceptibility
H.R., 2019. Artificial neural networks for flood susceptibility mapping in data-scarce using random-forest and boosted-tree models in Seoul metropolitan city, Korea.
urban areas. Spatial Modeling in GIS and R for Earth and Environmental Sciences. Geomat. Nat. Hazards Risk 8 (2), 1185–1203.
Elsevier 323-336. Li, X., Zhao, Z., Zhu, X., Wyatt, T., 2011. Covering models and optimization techniques for
Frigerio, I., Ventura, S., Strigaro, D., Mattavelli, M., De Amicis, M., Mugnano, S., Boffi, M., emergency response facility location and planning: a review. Math. Meth. Oper. Res.
2016. A GIS-based approach to identify the spatial variability of social vulnerability 74 (3), 281–310.
to seismic hazard in Italy. Appl. Geogr. 74, 12–22. Masood, M., Takeuchi, K., 2012. Assessment of flood hazard, vulnerability and risk of mid-
Ganguly, K.K., Nahar, N., Hossain, B.M., 2019. A machine learning-based prediction and eastern Dhaka using DEM and 1D hydrodynamic model. Nat. Hazards 61 (2),
analysis of flood affected households: A case study of floods in Bangladesh. Int. 757–770.
J. Disast. Risk Reduct. 34, 283–294. Masuya, A., Dewan, A., Corner, R.J., 2015. Population evacuation: evaluating spatial distri-
Ghenai, C., Albawab, M., Bettayeb, M., 2020. Sustainability indicators for renewable energy bution of flood shelters and vulnerable residential units in Dhaka with geographic in-
systems using multi-criteria decision-making model and extended SWARA/ARAS hy- formation systems. Nat. Hazards 78 (3), 1859–1882.
brid method. Renew. Energy 146, 580–597. Mestre, A.M., Oliveira, M.D., Barbosa-Póvoa, A.P., 2015. Location–allocation approaches for
Gnyawali, K.R., Zhang, Y., Wang, G., Miao, L., Pradhan, A.M.S., Adhikari, B.R., Xiao, L., 2019. hospital network planning under uncertainty. Eur. J. Oper. Res. 240 (3), 791–806.
Mapping the susceptibility of rainfall and earthquake triggered landslides along Mirchooli, F., Motevalli, A., Pourghasemi, H.R., Mohammadi, M., Bhattacharya, P.,
China–Nepal highways. Bull. Eng. Geol. Environ. 79 (2), 587–601. Maghsood, F.F., Tiefenbacher, J.P., 2019. How do data-mining models consider arsenic
Gu, W., Wang, X., McGregor, S.E., 2010. Optimization of preventive health care facility lo- contamination in sediments and variables importance? Environ. Monit. Assess. 191
cations. Int. J. Health Geogr. 9 (1), 17. (12), 777.
Haghizadeh, A., Siahkamari, S., Haghiabi, A.H., Rahmati, O., 2017. Forecasting flood-prone Mohammady, M., Pourghasemi, H.R., Amiri, M., 2019. Assessment of land subsidence sus-
areas using Shannon’s entropy model. J. Earth Syst. Sci. 126 (3), 39. ceptibility in Semnan plain (Iran): a comparison of support vector machine and
Haltas, I., Tayfur, G., Elci, S., 2016. Two-dimensional numerical modeling of flood wave weights of evidence data mining algorithms. Nat. Hazards 99 (2), 951–971.
propagation in an urban area due to Ürkmez dam-break, İzmir, Turkey. Nat. Hazards Mojaddadi, H., Pradhan, B., Nampak, H., Ahmad, N., Ghazali, A. H. b., 2017. Ensemble
81 (3), 2103–2119. machine-learning-based geospatial approach for flood risk assessment using multi-
Hateffard, F., Dolati, P., Heidari, A., Zolfaghari, A.A., 2019. Assessing the performance of de- sensor remote-sensing data and GIS. Geomat. Nat. Hazards Risk 8 (2), 1080–1102.
cision tree and neural network models in mapping soil properties. J. Mt. Sci. 16 (8), Polo, G., Acosta, C.M., Ferreira, F., Dias, R.A., 2015. Location-allocation and accessibility
1833–1847. models for improving the spatial planning of public health services. PLoS One 10
Hirabayashi, Y., Kanae, S., 2009. First estimate of the future global population at risk of (3), e0119190.
flooding. Hydrol. Res. Lett. 3, 6–9. Pourghasemi, H.R., Gayen, A., Edalat, M., Zarafshar, M., Tiefenbacher, J.P., 2019a. Is multi-
Holmes, J., Williams, F.B., Brown, L.A., 1972. Facility location under a maximum travel re- hazard mapping effective in assessing natural hazards and integrated watershed
striction: an example using day care facilities. Geogr. Anal. 4 (3), 258–266. management? Geosci. Front. 11 (4), 1203–1217.
Hong, H., Tsangaratos, P., Ilia, I., Liu, J., Zhu, A.X., Chen, W., 2018. Application of fuzzy Pourghasemi, H.R., Gayen, A., Panahi, M., Rezaie, F., Blaschke, T., 2019b. Multi-hazard
weight of evidence and data mining techniques in construction of flood susceptibility probability assessment and mapping in Iran. Sci. Total Environ. 692, 556–571.
map of Poyang County, China. Sci. Total Environ. 625, 575–588. Pourghasemi, H.R., Nitheshnirmal, S., Karimi, N., Collins, A., 2020. Gully erosion spatial
Hossain, S., 2013. Flood Damage and Risk Assessment Model in the Haor Basin of modelling: role of machine learning algorithms in selection of the best controlling
Bangladesh. Bangladesh University of Engineering and Technology, M.S. thesis. factors and modelling process. Geosci. Front. 11 (6), 2207–2219.
Hunt, A., Watkiss, P., 2011. Climate change impacts and adaptation in cities: a review of Pourghasemi, H.R., Kariminejad, N., Amiri, M., Edalat, M., Zarafshar, M., Blaschke, T., Cerda,
the literature. Clim. Chang. 104 (1), 13–49. A., 2020a. Assessing and mapping multi-hazard risk susceptibility using a machine
Hunt, E.B., Marin, J., Stone, P.J., 1966. Experiments in Induction. learning technique. Sci. Rep. 10 (1), 1–11.
Islam, K.N., 2011. Impacts of Urban Floods from Micro-Macro Level Perspectives: A Case Pourghasemi, H.R., Termeh, S.V.R., Kariminejad, N., Hong, H., Chen, W., 2020b. An assess-
Study of Bangladesh. LAP LAMBERT academic Publishing. ment of metaheuristic approaches for flood assessment. J. Hydrol. 582, 124536.
Islam, M., Sado, K., 2000a. Flood hazard assessment in Bangladesh using NOAA AVHRR Rahman, M., Ningsheng, C., Islam, M.M., Dewan, A., Iqbal, J., Washakh, R.M.A., Shufeng, T.,
data with geographical information system. Hydrol. Process. 14 (3), 605–620. 2019. Flood susceptibility assessment in Bangladesh using machine learning and
Islam, M.M., Sado, K., 2000b. Development of flood hazard maps of Bangladesh using multi-criteria decision analysis. Earth Syst. Environ. 3 (3), 585–601.
NOAA-AVHRR images with GIS. Hydrol. Sci. J. 45 (3), 337–355. Rahmati, O., Pourghasemi, H.R., Zeinivand, H., 2016. Flood susceptibility mapping using
Islam, M.M., Sado, K., 2002. Development priority map for flood countermeasures by re- frequency ratio and weights-of-evidence models in the Golastan Province, Iran.
mote sensing data with geographic information system. J. Hydrol. Eng. 7 (5), Geocart. Int. 31 (1), 42–70.
346–355. Reduction, I. S. F. D, 2004. Living with Risk: A Global Review of Disaster Reduction Initia-
Janizadeh, S., Avand, M., Jaafari, A., Phong, T.V., Bayat, M., Ahmadisharaf, E., Prakash, I., tives. United Nations Publications.
Pham, B.T., Lee, S., 2019. Prediction success of machine learning methods for flash Regmi, A.D., Yoshida, K., Pourghasemi, H.R., Dhital, M.R., Pradhan, B., 2014. Landslide sus-
flood susceptibility mapping in the Tafresh Watershed, Iran. Sustainability 11 (19), ceptibility mapping along Bhalubang—Shiwapur area of mid-Western Nepal using
5426. frequency ratio and conditional probability models. J. Mt. Sci. 11 (5), 1266–1285.
Jato-Espino, D., Lobo, A., Ascorbe-Salcedo, A., 2019. Urban flood risk mapping using an Rizeei, H.M., Pradhan, B., Saharkhiz, M.A., 2019. Allocation of emergency response centres
optimised additive weighting methodology based on open data. J. Flood Risk in response to pluvial flooding-prone demand points using integrated multiple layer
Manag. 12, e12533. perceptron and maximum coverage location problem models. Int. J. Disast. Risk
Javidi, M.M., Mansoury, S., 2017. Diagnosis of the disease using an ant colony gene selec- Reduct. 38, 101205.
tion method based on information gain ratio using fuzzy rough sets. J. Part. Sci. Romero, J.P., Ibeas, A., Moura, J.L., Benavente, J., Alonso, B., 2012. A simulation-
Technol. 3 (4), 175–186. optimization approach to design efficient systems of bike-sharing. Procedia Soc.
Kar, B., Hodgson, M.E., 2008. A GIS-based model to determine site suitability of emergency Behav. Sci. 54, 646–655.
evacuation shelters. Trans. GIS 12 (2), 227–248. Samanta, R.K., Bhunia, G.S., Shit, P.K., Pourghasemi, H.R., 2018. Flood susceptibility map-
Keršuliene, V., Zavadskas, E.K., Turskis, Z., 2010. Selection of rational dispute resolution ping using geospatial frequency ratio technique: a case study of Subarnarekha River
method by applying new step-wise weight assessment ratio analysis (SWARA). Basin, India. Model. Earth Syst. Environ. 4 (1), 395–408.
J. Bus. Econ. Manag. 11 (2), 243–258. Sarkar, D., Mondal, P., 2020. Flood vulnerability mapping using frequency ratio (FR)
Khan, S.N., Mir, K., Tahir, A., Awan, A., Nisa, Z.U., Gillani, S.A., 2018. Allocation of tutors and model: a case study on Kulik river basin, Indo-Bangladesh Barind region. Appl
study centers in distance learning using geospatial technologies. ISPRS Int. J. Geo Inf. 7 Water Sci 10 (1), 17.
(5), 185. Seejata, K., Yodying, A., Wongthadam, T., Mahavik, N., Tantanee, S., 2018. Assessment of
Kheir, R.B., Greve, M.H., Abdallah, C., Dalgaard, T., 2010. Spatial soil zinc content distribu- flood hazard areas using analytical hierarchy process over the lower Yom Basin,
tion from terrain parameters: a GIS-based decision-tree model in Lebanon. Environ. Sukhothai Province. Proc. Eng. 212, 340–347.
Pollut. 158 (2), 520–528. Sherali, H.D., Carter, T.B., Hobeika, A.G., 1991. A location-allocation model and algorithm
Khosravi, K., Nohani, E., Maroufinia, E., Pourghasemi, H.R., 2016. A GIS-based flood suscep- for evacuation planning under hurricane/flood conditions. Transp. Res. B Methodol.
tibility assessment and its mapping in Iran: a comparison between frequency ratio 25 (6), 439–452.
16
Siahkamari, S., Haghizadeh, A., Zeinivand, H., Tahmasebipour, N., Rahmati, O., 2018. Spa- Wang, Y., Fang, Z., Hong, H., 2019. Comparison of convolutional neural networks for land-
tial prediction of flood-susceptible areas using frequency ratio and maximum en- slide susceptibility mapping in Yanshan County, China. Sci. Total Environ. 666,
tropy models. Geocart. Int. 33 (9), 927–941. 975–993.
Silva, G., 2016. Event based flood inundation mapping under the impact of climate Whitfield, P., 2012. Floods in future climates: a review. J. Flood Risk Manag. 5 (4),
change: a case study in lower Kelani River Basin, Sri Lanka. Hydrol. Curr. Res. 7 (1), 336–365.
1000228. Xi, M., Ye, F., Yao, Z., Zhao, Q., 2013. A modified-median model for the emergency facilities
Tehrany, M.S., Kumar, L., 2018. The application of a Dempster–Shafer-based evidential be- location problem and its variable neighbourhood search-based algorithm. J. Appl.
lief function in flood susceptibility mapping and comparison with frequency ratio and Math. 2013, 1–10.
logistic regression methods. Environ. Earth Sci. 77 (13), 490. Xiong, J., Sun, M., Zhang, H., Cheng, W., Yang, Y., Sun, M., Cao, Y., Wang, J., 2019. Applica-
Tehrany, M.S., Pradhan, B., Jebur, M.N., 2015. Flood susceptibility analysis and its verifica- tion of the Levenburg–Marquardt back propagation neural network approach for
tion using a novel ensemble support vector machine and frequency ratio method. landslide risk assessments. Nat. Hazards Earth Syst. Sci. 19 (3), 629–653.
Stoch. Env. Res. Risk A. 29 (4), 1149–1165. Xu, W., Ma, Y., Zhao, X., Li, Y., Qin, L., Du, J., 2018. A comparison of scenario-based hybrid
Tehrany, M.S., Kumar, L., Shabani, F., 2019. A novel GIS-based ensemble technique for bilevel and multi-objective location-allocation models for earthquake emergency
flood susceptibility mapping using evidential belief function and support vector ma- shelters: a case study in the central area of Beijing, China. Int. J. Geogr. Inf. Sci. 32
chine: Brisbane, Australia. PeerJ 7, e7653. (2), 236–256.
Tingsanchali, T., Karim, M.F., 2005. Flood hazard and risk analysis in the southwest region Yesilnacar, E., Topal, T., 2005. Landslide susceptibility mapping: a comparison of logistic
of Bangladesh. Hydrol. Process. 19 (10), 2055–2069. regression and neural networks methods in a medium scale study, Hendek region
Tong, D., Murray, A.T., 2009. Maximising coverage of spatial demand for service. Pap. Reg. (Turkey). Eng. Geol. 79 (3–4), 251–266.
Sci. 88 (1), 85–97. Zavadskas, E.K., Turskis, Z., Ustinovichius, L., Shevchenko, G., 2010. Attributes weights
Wahab, A., Ludin, A.M., 2018. Flood vulnerability assessment using artificial neural net- determining peculiarities in multiple attribute decision making methods. Eng.
works in Muar Region, Johor Malaysia. IOP Conference Series: Earth and Environmen- Econ. 66 (1).
tal Science. 169, p. 012056. Zolfani, S.H., Yazdani, M., Zavadskas, E.K., 2018. An extended stepwise weight assessment
Wahlstrom, M., Guha-Sapir, D., 2015. The Human Cost of Weather-related Disasters ratio analysis (SWARA) method for improving criteria prioritization process. Soft.
1995–2015: Geneva. UNISDR, Switzerland. Comput. 22 (22), 7399–7405.
Wang, Y., Li, Z., Tang, Z., Zeng, G., 2011. A GIS-based spatial multi-criteria approach for
flood risk assessment in the Dongting Lake Region, Hunan, Central China. Water
Resour. Manag. 25 (13), 3465–3484.
17

2 Scopusresults

Uploaded by

Copyright:

Available Formats

2 Scopusresults

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2 Scopusresults

Uploaded by

Copyright:

Available Formats

Geoscience Frontiers 12 (2021) 101095

Contents lists available at ScienceDirect

journal homepage: www.elsevier.com/locate/gsf

Location-allocation modeling for emergency evacuation planning with

1. Introduction elevation. The ﬂooding has severe implications on livelihoods, food

Fig. 2. Schematic framework of the methodology.

Table 1 2.4. Spatial modeling

to check the performance of model. The number of hidden layers ranged

where, n is the number of factors, wk is the weighting coefﬁcients for

bility map by multiplying it with the ﬂood triggering layer.

∑xij ≤1 for all i

Tolerance VIF Tolerance VIF

1 Elevation 0.104 0.540 1.851 9 Drainage density 0.060 0.742 1.347

3.4. Preparation of ﬂood susceptibility maps

In the training of LM-BP model, the neural network was stopped

and map accuracy quantitatively. Yesilnacar and Topal (2005) men-

3.6. Preparation of location-allocation models

The ﬁrst step toward implementing the LAM is to locate existing

PMP MA PMP MA PMP MA PMP MA

Low 28,172 45.33 −20,509 −9973 −5496 −2998 −1633 −1077

existing EECs. Therefore, more EECs would be required to provide ac-

You might also like