You are on page 1of 30

Geocarto International

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tgei20

Flood susceptibility modeling in the urban


watershed of Guwahati using improved
metaheuristic-based ensemble machine learning
algorithms

Ishita Afreen Ahmed, Swapan Talukdar, Shahfahad, Ayesha Parvez, Mohd.


Rihan, Mirza Razi Imam Baig & Atiqur Rahman

To cite this article: Ishita Afreen Ahmed, Swapan Talukdar, Shahfahad, Ayesha Parvez, Mohd.
Rihan, Mirza Razi Imam Baig & Atiqur Rahman (2022): Flood susceptibility modeling in the
urban watershed of Guwahati using improved metaheuristic-based ensemble machine learning
algorithms, Geocarto International, DOI: 10.1080/10106049.2022.2066200

To link to this article: https://doi.org/10.1080/10106049.2022.2066200

Published online: 15 Jun 2022.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tgei20
GEOCARTO INTERNATIONAL
https://doi.org/10.1080/10106049.2022.2066200

Flood susceptibility modeling in the urban watershed of


Guwahati using improved metaheuristic-based ensemble
machine learning algorithms
Ishita Afreen Ahmeda, Swapan Talukdara , Shahfahada , Ayesha Parvezb,
Mohd. Rihana , Mirza Razi Imam Baiga and Atiqur Rahmana
a
Department of Geography, Faculty of Natural Sciences, Jamia Millia Islamia, New Delhi, India;
b
Department of Electrical Engineering and Computer Science, Henry Samueli School of Engineering,
University of California, Irvine, CA, USA

ABSTRACT ARTICLE HISTORY


The urban watershed of Guwahati is a highly flood-prone region Received 5 December 2021
and the fastest growing city situated on the bank of the Accepted 10 April 2022
Brahmaputra River. Therefore, this study aims to the urban flood
KEYWORDS
susceptibility mapping of Guwahati city using metaheuristic opti-
Flood susceptibility
mization algorithms integrated with random forest (RF) machine mapping; urban watershed;
learning algorithm. Further, the receiver operating characteristic metaheuristic optimization
(ROC) and multiple error measurements were applied to analyze algorithms; particle swarm
the performances of the models used. The result showed that optimization; machine
about one-third of the area of Guwahati city is under the high learning algorithms
and very high flood risk while nearly 50% area comes under low
and very low flood risk. The value of the area under curve (AUC)
of ROC was above 0.80 for all the integrated models applied.
However, the RF-bee colony (BCO) and the RF-based ant colony
(ACO) are the two best flood susceptibility models that performed
better in the analysis. The methodology adopted in the study is
cost and time effective and can be used for the flood
susceptibility modeling in other parts of the world. Further, the
findings of this study can useful in the flood mitigation and
planning process.

1. Introduction
Floods are the most disastrous and catastrophic natural hazards that affect almost 170
million people annually throughout the globe (Kazakis et al. 2015; Wang et al. 2019).
Over the past twenty years, the frequency of floods worldwide has increased by over 40%.
From 1995 to 2015, about 109 million people around the globe were affected by floods;
with damages of about 75 billion USD per year (Alfieri et al. 2017). The developing coun-
tries like India, China, Bangladesh and Brazil, are frequently hit by such natural calamities
(Lucas and Kibler 2016; Imamura 2022). In India, nearly 40 Mha of land area is suscep-
tible to the flood hazards (Tripathi et al. 2022). Indian floods, on average, have a devastat-
ing effect on around 7.5 million hectares of land, resulting in the loss of more than 1600

CONTACT Atiqur Rahman arahman2@jmi.ac.in


ß 2022 Informa UK Limited, trading as Taylor & Francis Group
2 I. A. AHMED ET AL.

lives each year (National Disaster Management Authority of India 2008). Both natural
and anthropogenic factors are responsible for causing flood events (Chang and Chen
2016; Nguyen et al. 2021), but changes in the world climatic conditions is one of the
main reasons behind variations in flood patterns and flood intensity and magnitude. The
major cause of flood in rural areas is the high streamflow of rivers, while in urban areas,
flood is mainly caused by an increase in the imperviousness of the surface and poor
drainage system (Rubinato et al. 2019). Due to continuous construction practices, the pre-
vention of urban flooding has become almost difficult, with a significant emphasis on
flood control aimed at mitigating the loss of life and property. In the past few decades,
numerous studies have been reported relating to the development of flood forecasting and
prediction models in rural and urban areas (Nkwunonwo et al. 2020).
One of the most significant issues which can enhance the sophistication of environ-
mental studies such as flood is the degree of urbanization (Ahmed et al. 2021). As a con-
sequence of tremendous growth in the construction of infrastructures and the resultant
reduction in the accommodating capacity of the subsurface for precipitation, the flood in
urban areas has been aggravating at a rapid rate. Urban flood disasters have been swelling
up rapidly in India over the past 30 years which severely affected the major cities in
India. Some of the most recent disastrous flood-events in India are Hyderabad flood
(2000), Ahmedabad flood (2001), Delhi flood (2002, 2003 and 2009), Chennai floods
(2004 and 2015), floods in Mumbai, Surat and Kolkata in 2005, 2006 and 2007, and
Guwahati flood in 2010 (Baghel 2016). Also, recurring floods in most parts of India are
very common, but states of northeastern India are struck worse by its ferocity
(NIDM 2011).
The urban watershed of Guwahati is a highly flood-prone region and a disaster hotspot
located on the Brahmaputra River. It is the largest urban and financial center of north-
eastern India, with an important route linking the region and the mainland (Sarmah et al.
2020). After every average to hefty shower, the urban watershed experiences inundation
and waterlogging problems all over the city. Moreover, the drainage system of Guwahati
is of poor quality and out of date as it cannot drain out floodwater during the monsoonal
rainfall (Sahoo and Sreeja 2017). Both natural and flash floods have been experienced in
the urban watershed of Guwahati over the years (Sarmah et al. 2020). Floods in the
Brahmaputra River are caused by extreme precipitation in the low-lying areas near the
wetlands and reservoirs, but flash floods have become a regular yearly event, primarily in
the city center (Kashyap and Mahanta 2018). Hence, the studies on flood events and map-
ping flood susceptibility zones are gaining importance in the urban watershed of
Guwahati for policy-making due to its increasing urban flooding intensity and recurring
pattern on an annual basis.
Lately, techniques and methods such as machine learning have gained more recogni-
tion from researchers for predicting flood susceptibility and creating models all over the
globe (Chen et al. 2019; Bui et al. 2020). Artificial neural networks (Falah et al. 2019;
Pham et al. 2021), support vector machines (Pandey et al. 2021), random forests (Chen
et al. 2020; Talukdar et al. 2020) and decision trees (Moghaddam et al. 2019; Nhu et al.
2020) are the most widely used machine learning methods used to forecast regions at risk
of flooding. But, machine learning modeling of flood susceptibility poses numerous chal-
lenges, such as selecting the right methods of modeling optimization from a vast range of
methods, and each approach delivers distinct results (Shafizadeh-Moghadam et al. 2018;
Costache and Tien Bui 2020). Many methods have been identified around the globe to
create flood susceptibility models (Khosravi et al. 2016; Al-Abadi 2018) using geological
and geomorphological parameters. Frequency ratio (FR) (Shafapour Tehrany et al. 2019),
GEOCARTO INTERNATIONAL 3

analytical hierarchy process (AHP) (Bahrami et al. 2021), logistic regression (LR) (Ali
et al. 2020), artificial neural network (ANN) (Bui et al. 2020; Yu and Chen 2020), decision
tree (DT) (Chen et al. 2020), adaptive neuro-fuzzy inference system (ANFIS) (Moayedi
et al. 2020), ANFIS-genetic algorithm (GA) (Hong et al. 2018), bagging–LMT hybrid
model (Talukdar et al. 2020) and random forest (Sun et al. 2020) and ANFIS-particle
swarm optimization (PSO) (Termeh et al. 2018) are methods with least agreement yet
they concern the assessment of natural disasters, such as flood susceptibility, for selecting
the most appropriate model. Also, due to changes in the geographical factors such as.,
topography, landscape setting and hydrology, of a specific study area, there is a variation
in the results of each optimization algorithm (Wang et al. 2020). Therefore, a number of
models must first be applied to a given area, and one can be chosen for use only after an
evaluation of the predictive superiority of the models.
In recent years, metaheuristic models such as Ant Colony Optimization (ACO)
(Termeh et al. 2021), PSO (Band et al. 2020), Tabu Search (Chou et al. 2021), Harmony
Search (Dodangeh et al. 2020), Cuckoo Search (Khan et al. 2020) and Bee Colony
Optimization (BCO) (Farfan and Cea 2021), are mostly used for the assessment of natural
disasters and hazards amid several machine learning algorithms for their competent calcu-
lation as well as computational efficiency. Many researchers have also utilized metaheuris-
tic optimization methods with machine learning models to forecast the susceptibility of
landslides, earthquakes, forest fires, etc. (Table 1). The results thus obtained indicate that
such an integration technique can be very effective in obtaining accurate information
about predicting any natural disasters and other components of the environment (Pham
et al. 2016).
Therefore, the present study uses metaheuristic optimization, including PSO, ACO,
bee colony optimization (BCO) and grey wolf optimization (GWO) integrated with
an ensembled-hybrid machine learning algorithm, random forest, for spatial prediction
of flood susceptibility in the urban watershed of Guwahati, Assam. Although numer-
ous machine learning algorithms and optimizations have been used in modelling glo-
bal flood events, the current optimization methods have not previously been explored
for flood event research. Thus, the present study aims to integrate the metaheuristic
optimization algorithms, such as ant colony, bee colony, grey wolf and particle
swarm, with machine learning algorithm for predicting urban flooding and it evaluate
the flood factors and create flood susceptibility maps in the Guwahati
urban watershed.

2. Study area
Guwahati’s urban watershed was chosen as the study area because it is the fastest-growing
city located on both banks of the Brahmaputra River. Guwahati is situated in the Kamrup
(Metro) district of Assam (Figure 1). Figure 1 depicts the total number of flood and non-
flood points considered for the survey, as well as three meteorological stations. The
Guwahati Municipal Corporation (GMC), the North Guwahati Town Committee (NGTC)
district, the Amingaon Census Town (ACT) and 21 revenue villages form the urban
watershed of Guwahati and are recognized as the Guwahati Metropolitan Area (GMA). It
occupies an area of 340.67 km2 in total., Guwahati’s overall population grew from 500,000
in 1991 to 814,575 in 2001 and 963,429 in 2011 (Pawe and Saikia 2020). The region is
dominated by an undulating topographical landscape with an elevation ranging from 50
m to 55 m above sea level (AMSL). By spreading between Lokpriya Gopinath Bordoloi
(LGB) International Airport in the west and Narengi in the east with an area of around
4 I. A. AHMED ET AL.

Table 1. Examples of previous researches using metaheuristic optimization models with machine learning models
for predicting susceptibility.
S. No Authors Study area Method Findings
1. Bui et al. (2016) Tuong Duong The neural fuzzy inference The results show that
district in system was used to create the proposed
Central Vietnam an initial flood metaheuristic
susceptibility model and optimization for the
then the model was flood susceptibility
optimized using two model outperforms
metaheuristic algorithms, the above
Evolutionary Genetic and benchmark models;
Particle Swarm also, the
Optimization metaheuristic
optimization for
flood susceptibility is
a new alternative
tool that should be
used in flood
susceptibility
mapping
2. Tien et al. (2017) Lao Cai A new hybrid intelligent Performance of the
area, Vietnam method based on least LSSVM-BC model
squares support vector shows that the
machines (LSSVM) and prediction power of
artificial bee colony (ABC) the model is good
optimization, with the area under
the
curve (AUC) ¼ 0.900
3. Termeh et al. (2018) Jahrom Township in combination of adaptive The ensemble of ANFIS-
Fars Province, Iran neuro-fuzzy inference PSO was introduced
systems (ANFIS) with as the premier
different metaheuristics model in the
algorithms such as ant study area.
colony optimization
(ACO), genetic
algorithm (GA) and particle
swarm optimization (PSO)
and comparing
their accuracy
4. Ahmadlou The southern part of Adaptive neuro-fuzzy ANFIS-BBO had the
et al. (2019) the Golestan inference system (ANFIS), highest accuracy in
Province, Iran with two heuristic-based comparison with the
computation methods ANFIS and ANFIS-BA
namely biogeography- models in
based optimization (BBO) flood modeling
and BAT algorithm (BA)
with GIS
5. Niyomubyeyi Kigali, Rwanda Multi-Objective Artificial Bee The proposed MOABC
et al. (2019) Colony (MOABC) algorithm, outperforms the
modified to provide a current methods
better solution to the both in terms of
evacuation problem computational time
combined with random and better solutions
swap and random insertion with minimum
methods for neighborhood fitness values
search, the two-point
crossover operator, and the
Pareto-based method.
6. Panahi et al. (2020) Icheon Township, Hybrid Machin learning and The SVR–GWO and
South Korea metaheuristic algorithms, SVR–Bee models
that is, the bee algorithm were the most
(Bee), the adaptive neuro- predictive models in
fuzzy inference system terms of constructing
(ANFIS), support vector the exceptionally
(continued)
GEOCARTO INTERNATIONAL 5

Table 1. Continued.
S. No Authors Study area Method Findings
regression (SVR) and the focused landslide
grey wolf optimizer (GWO) susceptibility map.
and to compare their
predictive accuracy
7. Zhang et al. (2020) Changsha city, China Application of machine The proposed hybrid
learning (ML) algorithms, intelligent algorithm
back-propagation (BPNN), integrated with
general regression neural metaheuristic
network (GRNN), extreme algorithm particle
learning machine (ELM), swarm optimization
support vector machine (PSO) and ML can
(SVM) and random forest effectively determine
(RF) in predicting the global optimum
tunneling-induced hyper-parameters of
settlement combined with ML algorithms.
metaheuristic algorithms,
particle swarm
optimization (PSO)
8. Dodangeh Haraz-Neka Novel hybrid intelligence The results showed that
et al. (2020) watershed in the models based on the meta- the SVR model had
coastal margins optimization of the support superior performance
of the vector regression (SVR) and (AUC ¼ 0.70–0.75,
Caspian Sea group method of data RMSE ¼ 0.29–0.36,
handling (GMDH) using MSE ¼ 0.08–0.13).
different metaheuristic Both the GA and HS
algorithms, i.e. the genetic remarkably improved
algorithm (GA) and the SVR and GMDH
harmony search (HS) performance, with
the SVR-GA model
performing the best.
9. Arora et al. (2021) Middle Ganga Plain Quantitatively test and The accuracy was found
(MGP), India compare novel advanced- to be highest for
machine learning ANFIS-GA during
algorithms with an training (0.886) and
advanced novel hybrid validation (0.883).
model Adaptive Neuro- Better performance
Fuzzy Inference System of ANIFS-GA than
(ANFIS) and three the individual
metaheuristic models- models as well as
based ensembles with some ensemble
ANFIS namely ANFIS-GA models suggests and
(Genetic Algorithm), ANFIS- warrants further
DE (Differential Evolution) study in this topo
and ANFIS-PSO (Particle climatic environment
Swarm Optimization)
10. Hagighi et al. (2021) Pole-Doab Employed four machine Assessments of model
watershed, Iran learning algorithms performance
[Support Vector Machine indicated that DA
(SVM), Multivariate had the highest
Adaptive Regression accuracy and
Splines (MARS), Generalized efficiency, with the
Linear Model (GLM) and greatest learning and
Dragonfly Algorithm (DA)] prediction power in
for land degradation (LD) LD risk mapping
risk mapping

45 km, the urban watershed of Guwahati is more prevalent in the east-west pattern (Pawe
and Saikia 2018).
A mild and humid climate with a temperature range between 12  C and 14  C in the
winter and 30–37  C during the summer season characterizes the study area. During the
6 I. A. AHMED ET AL.

Figure 1. Location of the study area.

monsoon season, the maximum precipitation is observed, i.e. from June to September
with an average precipitation of 200–350 mm/year. Except for the winter seasons of
December and January, the temperature of the study area normally fluctuates due to a
high level of humidity during the year. Low-Precambrian residual hills and inselbergs
along with elongated and linear low-lying valley fills, marshes and wetlands are largely
part of the urban watershed of Guwahati (Yadav and Barua 2018).

3. Database and methodology


3.1. Database
The urban watershed of Guwahati experiences repeated flood events each year; hence, the
historical flood inventories were prepared in this study primarily based on the field survey
and the understanding of local citizens. A total of thirteen flood influencing parameters
were used to create the flood susceptibility model. Satellite image Landsat-8 (OLI/TIRS)
was collected from the United States Geological Survey (USGS) website (https://earth-
explorer.usgs.gov/) to plan land use/land cover maps. The topographical, as well as hydro-
logical factors of the study area were derived from the SRTM (spatial resolution 30 m)
map. The precipitation data of the study area were collected from the Regional
Meteorological Centre, Guwahati, Assam. Detailed information about the database of this
study is presented in Table 2.
GEOCARTO INTERNATIONAL 7

Table 2. Details of the data used.


Extractable
S. No. Data source Year information
1. Google Earth 2019 City Map
Path Row
2. Satellite Data Landsat Optical and Thermal
(USGS 8 (OLI þ TIRS) 137 42 2020 Data (Land use/
Earth cover, Vegetation,
Explorer) Temperature, etc.)
SRTM 137 42 2000 Digital
Elevation Model
3. Indian Meteorological Department, Regional Meteorological Climatic data such
1991–2019 as rainfall,
temperature,
relative humidity
(Lokpriya
Gopinath
Bordoloi
International
Airport
Meteorological
Station)
4. Flood Inventory Map 2001–2020 Historical
flood records

3.2. Methodology
The methodological flow chart of the present study can be seen in Figure 2. The flowchart
demonstrates the outline of the whole analysis, which includes the development of the
flood inventory map as well as the generation of flood influencing factors, the ROC valid-
ity assessment of flood influencing factors and the multicollinearity measure. In addition,
the development of flood susceptibility models using machine learning algorithms,
Random Forest with Ant Search, Bee Colony, Grey Wolf and Particle Swarm metaheuris-
tic optimization algorithms methods were used for the current research.

3.2.1. Preparing flood inventory map


Employing multiple analyses, a flood inventory map composed of past flood data of an
area can be used to forecast possible flood events (Khosravi et al. 2018; El-Magd 2022).
Utilizing 292 points, the flood inventory map of the urban watershed of Guwahati was
prepared, and the positions of flood incidents were checked from flood reports by
detailed field surveys and Google Earth with the aid of GPS. Later, 80% (234 points)
and 40% (58 points) of the 292 randomly obtained flood points were divided into cate-
gories to train and test flood-susceptible models. Since the flood inventory map was div-
ided into two groups, flood points and non-flood points, modelling an area’s flood
vulnerability is considered a binary classification (El-Haddad et al. 2021; El-Magd et al.
2021). Based on the training datasets, we extracted the information from flood condi-
tioning parameters using extract values to point tool. Then, these datasets were used for
implementing the metaheuristic-based RF models. After generating flood susceptible
models, we extracted the prediction results from the flood susceptible models based on
the testing sample location. Then, the data were compared with the ground truth using
validation measures.
8 I. A. AHMED ET AL.

3.2.2. Method for preparing flood conditioning factors


The mapping of an area’s flood susceptibility is typically loaded with complexities and
inclusiveness, as various landscape features and hydrological factors are required in geo-
spatial format. Thirteen flood affecting parameters focused on former studies and evalua-
tions on flood susceptibility modeling were chosen and prepared in the current analysis
(Janizadeh et al. 2019; Arabameri et al. 2020). Based on the previous literature review on
the selection of the flood conditioning parameters, we selected thirteen parameters, such
as elevation, curvature, aspect, slope, flow direction, flow accumulation, topographic
roughness index (TRI), topographic wetness index (TWI), stream power index (SPI),
Land use and land cover, distance from the river, distance from the road and rainfall
(Tehrany et al. 2019; Kalantar et al. 2021; Panahi et al. 2021). Because of their importance
in the construction of precise mapping of potential flood regions, these parameters have
been used by numerous scholars for flood conditioning in previous studies. The analysis
and visualization of data were performed in a GIS environment with raster format data
and a spatial resolution of 30 m (Table 2).

3.2.2.1. Elevation. The Digital Elevation Model (DEM) obtained from the Shuttle Radar
Topography Mission (SRTM), has a 30 m resolution and is provided by NASA (https://
earthexplorer.usgs.gov) was used to generate the elevation model of the urban watershed
of Guwahati (Figure 3a). Plain areas or areas with lower elevation are usually more sus-
ceptible to floods. Therefore, it is very important to understand the topography and
derivative features responsible for floods in an area and assess the flood susceptibility.

3.2.2.2. Slope. Slope is a significant parameter in the study of flood mapping as it gives a
clear view of the degree of the slope angle. It greatly influences the velocity and the direc-
tion of movement of the floodwater. Here, the degree of slope angle ranges from 0 to
46.5 degrees implying gentle to very steep slopes (Figure 3b).

3.2.2.3. Aspect. Aspect is another parameter that affects the flooding scenarios of an area.
It provides a clear image of the direction of the slope inclination, thus, signifying the dir-
ection of the movement of floodwater (Figure 3c).

3.2.2.4. Curvature. Also, the present study used the profile curvature (Figure 3d) as it has
a great impact on the water flow velocity draining the surface. The curvature of the study
area was calculated using the Spatial Analyst Tool in ArcGIS 10.3software. The curvature
of an area is usually classified into three classes, i.e. concave, flat and convex. It is believed
that convex surfaces are more prone to runoff and are highly susceptible to floods
(Il’inskii and Yakimov 1987).

3.2.2.5. Flow accumulation. Flow accumulation is also used in the study as an important
parameter (Figure 3e). It computes the flow of water accumulated as the compiled weight
of all cells, and hence, flowing into each downslope cell in the output raster (Zhang
et al. 2017).

3.2.2.6. Flow direction. The direction of the steepest slope is often used to calculate the
direction of flow from each cell, which can also be considered a maximum drop. From a
digital surface model, the flow direction map is created. The final map shows eight appro-
priate output directions that connect to the eight neighboring cells into which flow could
GEOCARTO INTERNATIONAL 9

Table 3. Parameters of RF-ant colony model optimization.


RF-ant colony optimization
Start set no attributes
Population size 20
Number of iterations 20
Mutation type bit-off
Accelerate type accelerate
Objective type merits
Chaotic population type normal
Chaotic parameter type normal
Logistic map coefficient : 4.0
Mutation probability 0.01
Initial tau 0.1
Evaporation rate 0.9
Pheromone rate 2.0
Heuristic rate 0.7
Report frequency 20
Seed 1

Table 4. Parameters of RF-bee colony model optimization.


RF-bee colony optimization
Start set no attributes
Population size 30
Number of iterations 50
Mutation type bit-flip
Accelerate type normal
Objective type multiple objectives
Chaotic population type normal
Chaotic parameter type normal
Logistic map coefficient 4.0
Mutation probability 0.01
Radius Damp 0.98
Radius Mutation 0.8
Report frequency 30
Seed 1

Table 5. Parameters of RF-particle swarm model optimization.


RF-particle swarm optimization
Start set no attributes
Population size 20
Number of iterations 20
Mutation type bit-flip
Mutation probability 0.01
Inertia weight 0.33
Social weight 0.33
Individual weight 0.34
Report frequency 20
Seed: 1

pass (Figure 3f). This method is commonly referred to as the eight-direction (D8) flow
model and incorporates an approach proposed by Jenson and Domingue (1988).

3.2.2.7. Distance from river and distance from roads. The layers of distance from the
river (Figure 3g) and distance from the roads (Figure 3h) were obtained with the help of
Euclidean distance of the Spatial Analyst tool in ArcMap 10.3. The possibility of floods in
an area rises with a decline in distance from the river while possibilities fall with the
increase in distance from the river (Shahabi et al. 2021).
10 I. A. AHMED ET AL.

Figure 2. Methodological flowchart for flood susceptibility analysis.

3.2.2.8. Rainfall. One of the most influencing factors of flood has been identified as rain-
fall. The intensity of rainfall in an area for even a short period can cause disastrous flood
scenarios (Ali et al. 2020). The long-term precipitation data for the urban watershed of
Guwahati was collected from the Regional Meteorological Centre, Guwahati, Assam. The
kriging technique used in the study has been suggested to plot a smaller number of obser-
vations. Thus, the technique was employed for the study area as the rainfall data obtained
was only for three meteorological stations (Figure 3i).

3.2.2.9. Land use land cover. The land use land cover (LULC) of a region also has a sig-
nificant effect on the transport of surface runoff and sediments, both directly and indir-
ectly (Zhang et al. 2010). Compared to the forest and open areas, flooding in extremely
concrete settlement areas is more common. Typically, built-up lands do not allow water
to reach the soil and thereby obstruct an area’s surface runoff system (Costache and Tien
Bui 2020), whereas the forested and open surfaces are distinguished by a permeable envir-
onment that poses no barrier to water flow (Yin et al. 2017). The LULC map was estab-
lished in this analysis using the Landsat 8 (OLI/TIRS) dataset. In the present study, we
used maximum likelihood classifier for classifying the land use. To do so, we obtained the
training sample in terms of ground truth from Google Earth images and field survey. We
GEOCARTO INTERNATIONAL 11

Figure 3. Data layers for flood susceptibility in the urban watershed of Guwahati (a) elevation (b) slope (c) aspect, (d)
curvature, (e) flow accumulation, (f) flow direction, (g) distance from river, (h) distance from roads, (i) rainfall, (j) land
use/land cover (LULC), (k) topographic roughness index (TRI), (l) topographic wetness index (TWI), (m) stream power
index (SPI).

used 100 samples as training the model and rest of the 20 samples were used for validat-
ing the generated LULC. A total of eight LULC classes have been identified in this study,
i.e. water bodies, vegetation cover, agricultural croplands, grassland, bare/fallow/open
land, built-up area, river beds and river Brahmaputra (Figure 3j). In the present study, we
extracted the LULC data from the corresponding location of the testing or validating sam-
ple. Then, we applied kappa coefficient for validating the model. The results of accuracy
assessment show that the LULC had achieved 84% of overall accuracy. Therefore, it can
be considered as the satisfactory LULC model, which can be used for further research.

3.2.2.10. Terrain roughness index. Floods in an area are also estimated using the terrain
roughness/ruggedness index (TRI) which largely depends on the local topography of a
basin. The high value of TRI depicts either no or low flood while a lower value of TRI is
12 I. A. AHMED ET AL.

Figure 3. Continued.

always associated with higher floods (Tehrany et al. 2019). In this study, with the help of
the stretch format, the TRI map was prepared, with values ranging between 0.11 and 0.88
(Figure 3k).

3.2.2.11. Topographic wetness index. Apart from TRI, the TWI is also included as a par-
ameter in the study. TWI of an area depicts the wetness of a watershed by spatial differ-
ence. It was first proposed by Beven and Kirkby (1979). In other words, the TWI
represents the spatial variation of wetness of a river basin (Meles et al. 2020). The TWI
has been calculated using Eq. (1).
GEOCARTO INTERNATIONAL 13

 
As
TWI ¼ Ln [1]
tanb
where, As signifies the clear catchment area (m2 =m) and b denotes the slope gradient (in
degrees). In this study, the value of TWI ranges between 8.24 and 2.35 (Figure 3l).

3.2.2.12. Stream power index. The SPI of an area denotes the erosional power of the flow-
ing water. Generally, the SPI of a region has a significant effect on its fluvial systems
(Tehrany et al. 2014). The capacity to transport sediments and the erodibility of a river
from its bed is known as the SPI (Chen et al. 2020). The SPI was calculated by following
Moore et al. (1991) (Eq. (2))
SPI ¼ As tanb [2]
where As and the slope gradient representing the specific catchment area is indicated by b
(radians) (Figure 3m).

3.2.3. Method for flood influencing factors using multicollinearity test


For both likelihood models, several multicollinearity tests are used worldwide to evaluate
and measure influencing variables. Some of them are Pearson’s correlation coefficients,
variance decomposition proportions, conditional index, variance inflation factors (VIF)
and tolerances (Javidan et al. 2020). In this study, we have utilized the tolerances and VIF
tests to quantify the thirteen flood influencing factors (Dormann et al. 2013). We have
used the tolerances and VIF measurements in this analysis to measure the thirteen varia-
bles influencing floods. The VIF test estimates the strictness of multicollinearity with
respect to ordinary least square regression in statistical analysis. It is expressed by an
index that measures how much multicollinearity increases the variance of an approximate
regression coefficient. Now, if the VIF value of each flood influencing factor is greater
than 9, there is an issue of multicollinearity in that particular factor/factors (Duque and
Aquino 2019). The tolerance value less than 0.1 also indicate a multicollinearity question
(Khosravi et al. 2018). It is also strongly recommended to remove such influencing varia-
bles with a VIF value greater than 9 and tolerance lower than 0.1 before modeling.

3.2.4. Methods for flood susceptibility modeling


3.2.4.1. Machine learning algorithm. In a machine learning algorithm, a random-forest
model is an ensemble-learning technique that constructs a multitude of decision trees.
This model comprises an ensemble of simple classification and regression tree predictors
that explain the spatial relationship between the occurrence of a hazard and its associated
variables (Breiman 2001). The RF model now consists of multiple tree predictors, such
that each tree is distributed among all trees based on their randomly selected vector val-
ues in the forest (Cutler and Stevens 2006). The basic idea of the random forest model is
to create an immense number of different decision tree models for training. Every deci-
sion tree independently predicts the classification of the sample (for classification algo-
rithm). The mode of the sample classification is the final output. By creating unrelated
training sets to reduce the model’s variance, the efficiency of the random forest can
be increased.
In this study, to clarify the spatial relationship between flood frequency and its influ-
encing factors, the RF model is joined with some other algorithms. Several rules depend-
ent on the response variable are laid down in the training of a regression tree to separate
observations before the resulting predictions have a minimum amount of node impurity.
14 I. A. AHMED ET AL.

As the sum of the square deviation between the expected and actual value Breiman (1984)
identified node impurity for regression trees. The rules of the regression tree are a group
of linear divisions of observation data that combine to create a nonlinear surface
for judgment.
One of the key issues with regression trees is that they are susceptible to outfitting the
training datasets, and thus, behave incompetently when unseen input is presented
(Talukdar et al. 2020). RF is an approach that seeks to overcome this limitation. In the
RF algorithm, a portion of the input records and predictor variables are randomly chosen
as input to the testing while an individual regression tree is trained. This procedure is
replicated with the number of regression trees defined by the modeler, since each data set
is trained on a randomly chosen subset of records and input variables while
constructing a group of regression trees. An RF model shapes this community of regres-
sion trees.

3.2.4.2. Metaheuristic optimization algorithms. Ant colony optimization model. The Ant
algorithm was first developed to solve optimization problems such as traveling salesman
as a multicriteria approach (Dorigo and Blum 2005). The algorithms of Ant have pro-
gressed from mutual intelligence. The basic premise is based on the theory that the easy
and minimal local connections of members or population groups can contribute to clever
group activity. The main benefits of collective intelligence may be the absence of central-
ized authority; the transfers of dispersed creatures, and speed and simultaneous efficiency.
In the ACO algorithm, by leaving a trail of different pheromones along the path, ants
support one another to choose the shortest path. This is the method of discovering the
right path to the best answer and learning away from the model (based on the relation-
ship between the training set and conditioning variables) and testing the route selected
(test set) through a back-and-forth experiment.
Particle swarm optimization model. Over the past decades, optimization approaches based
on particle intelligence techniques have been considerably improved. This technique’s key
property is that it models the behavior of natural organisms such as birds, fish and ants.
The high intensity and simplicity for many problems is the biggest dominance of PSO
over traditional optimization approaches (Kennedy and Eberhart 1997). In comparison to
evolutionary approaches, the location of particles in the PSO algorithm depends on their
velocity parameters (step motor).
Particles travel in the search space in this model, and the particle velocity is dynamic-
ally modified according to the background of behavior before and after moves so that
each particle needs to go to a better spot (Bui et al. 2018). All particles have a utility func-
tion and a velocity vector symbolizing the particle’s orientation and phase, respectively
(Clerc and Kennedy 2002). The PSO is used for both continuous and discrete nonlin-
ear concerns.
Bee colony algorithm. A novel swarm-based optimization technique that can effectively
solve global optimization problems is the bee colony algorithm (Karaboga and Basturk
2008). The algorithm was influenced by the traditional foraging behavior of honeybees
who, because of its simplicity, versatility and desirable efficiency, attracted broad atten-
tion from the academic community (Cheng and Hoang 2015). Three types of honeybees,
including employed bees, onlooker bees and scout bees, are taken into account in the
optimization methodology of the hive colony to perform the search method. In general,
an iteration of the optimization algorithm for the bee colony requires four steps: food
supply initialization, employed bee phase hired, bee phase onlooker and bee
phase scout:
GEOCARTO INTERNATIONAL 15

Table 6. Parameters of RF-grey wolf model optimization.


RF-grey wolf optimization
Start set no attributes
Population size 30
Number of iterations 30
Mutation type bit-flip
Accelerate type normal
Objective type multiple-objective
Chaotic population type attribute number
Chaotic parameter type chaotic parameter
Logistic map coefficient 4.0
Mutation probability 0.01
Beta zero 0.33
Absorption coefficient 0.001
Report frequency 20
Seed 1

i. Food source initialization: by randomly generating an NFS (number of food source)


location, the bee colony algorithm begins the searching process. One artificial bee is
permitted to attack one location of the food supply.
ii. Step of the employed bee: each bee employed selects a new location in the feasible
space of the food source depending on the vicinity of the food source previ-
ously selected.
iii. Onlooker bee phase: employed bees return to their hive and share with onlooker
bees knowledge about the food supply. Based on the mutual data, depending on the
probability value pi correlated with that food source, the onlooker bee selects a
food source.
iv. Scout bee phase: if, after a specified number of successive iterations, the fitness of a
food source cannot help to increase the consistency of the food, the food source is
discarded by the bee and replaced with a new food source.
Grey wolf optimizer model. Another metaheuristic algorithm is the grey wolf optimizer
(GWO) algorithm, developed from the hunting actions of the grey wolves and the essence
of the social hierarchy. The Grey wolves exist in packs with a rigid system of social super-
iority and mimic the hierarchy of leadership. The grey wolves have separate classes for
events, including prey hunting. There are four levels of the social dominant hierarchy of
the grey wolves: alpha (a), beta (b), delta (d) and omega (x). As decision-making leaders,
the alpha-wolves are at the top of the social chain, while other wolves follow them. The
b-wolves support and design the rulers. Situated at the next step, the d-wolves follow the
alpha- and beta-wolves. Finally, the x-wolves have to submit to all of them. Finally, to
both of them, the al-wolves have to submit. Via the set of random candidate solutions,
the GWO methodology has an optimization mechanism close to other metaheuristic algo-
rithms, so it is designed as a statistical model for grey wolves. In fact, alpha, beta and
alpha are the fittest, second optimal, third optimal and the remaining solutions,
respectively.

a. The encirclement of prey: In the first step, while hunting, the grey wolves threaten
and encircle the prey. The parameter D calculates the distance from the prey to the
grey wolf.
b. Hunting: The hunting behavior is guided by alpha, beta and delta, respectively, after
the stage of encircling the prey, as they have compressive knowledge about the loca-
tion of the prey.
16 I. A. AHMED ET AL.

c. Prey attacking (exploitation): When the prey ceases running, the act of hunting ends.
The Grey wolves are watching the prey and pursuing it. In the GWO algorithm, the
pursuing prey is called the exploration phase. In this method, the parameters alpha,
b and d have the responsibility for guidance roles.

3.2.5. Integration process


The RF is used in this work to simulate the nonlinear connection between flood condi-
tioning parameters as inputs and flood prediction as output. The metaheuristic algorithms
were used to tweak the RF hyper-parameters in order to achieve expert performance. In
general, various parameters in RF, such as the maximum depth of a tree, the minimum
number of samples in a node, and the number of classification and regression trees, must
be tuned to achieve optimum RF performance. A viable alternative is to use an appropri-
ate optimization technique, such as different heuristic algorithms, to fine-tune
these parameters.
In this work, we use ACO, PSO, Bee colony and GWO, which are metaheuristic algo-
rithms that simulate the important parameters of the RF model to increase accuracy.
Metaheuristic algorithms use agents to find the best answer, which is referred to as popu-
lation search. As a result, metaheuristic algorithms must first be optimized. In this work,
we improved four algorithms six times via trial-and-error process. Following that, we
found error levels that were almost identical in terms of convergence after fifth trial. As a
result, we preferred the metaheuristic algorithms’ optimum values after fifth trial. Tables
2–5 give the optimum values of four algorithms. After improving the metaheuristic tech-
niques, these were employed to fine-tune the random forest model’s hyper-parameters. It
should be noted that the four metaheuristic models vary in their properties; hence the
optimum value will differ. However, we used four metaheuristic techniques to tune the
RF model’s parameters. Also, we tried to keep the values of the hyper-parameters of the
RF model same for all models to compare the performance. The best hyper-parameters
were found to be n-estimator: 914, max depth: 16, min sample split: 2, min sample leaf:
1. However, we also provide the process of integration in a schematic form.
The BC optimization relationships (i.e. employed, onlookers and scout bees) find and
compare different positions (food sources) in the following, and aim to substitute the ini-
tial RF parameters with the most suitable alternatives. This procedure is carried out in
order to achieve the best solution by defining an objective function mean absolute
error (MSE).
Parameters initialization of PSO algorithm! training and testing of machine learning
algorithm with the initialized parameters! calculation of fitness function! fitness value
of each swarm of particle in reference to local and global best values! updating the vel-
ocity and position of each swarm of particle accordingly! reaching maximum number of
iteration? if not reached, starting again from the second stage! if it reached the max-
imum number of iteration, that would be the optimal parameters for the machine learn-
ing algorithms.
Parameter’s initialization of GW optimization! initializes the random position of ‘n’
grey wolf in ‘d’ dimension! find the fitness value of wolf! stopping criteria met? if not
met, starting again from the second stage! update the position of wolf! calculate the
fitness of the wolf! find the value of alpha, beta and delta’s position! stopping criteria
met? If met, then starts with the training of machine learning algorithms! evaluate the
accuracy! meeting stopping criteria? If yes, then, the optimal model is obtained
(Table 6). If not meet the stopping criteria, then starts again with the second steps of GW
optimization.
GEOCARTO INTERNATIONAL 17

3.2.6. Validation process


Finally, each model’s performance and accuracy were measured in two stages.
Furthermore, by any objective function (’forecasting mistake’), such as root means square
error (RMSE), correlation coefficient, mean square error (MSE) and Spearman’s rho, con-
firmation during the training phase was integrated. The differences in the observed, as
well as expected values of the datasets, are primarily used to find RMSE and MAE. While
various error indices are used to determine the models’ predictive potential, MAE and
RMSE are the most commonly used methods for the standard metric measurement of
model errors in geosciences. Nonparametric methods used to calculate the force of inter-
action between two variables are the correlation coefficient and Spearman’s Rho, where
the value r ¼ 1 indicates a perfect positive correlation and the value r ¼ 1 means a
perfect negative correlation.
Second, we used the receiver operating characteristic curve (ROC) as well as the area
under the ROC curve (AUC) for further examination of the models’ predictive capability
for the validation of the testing process. For each model, the parameters giving the highest
precision were then selected as the optimal parameters and described in Table 7. For each
model, the ROC curve is shown in Figure 5. Either a positive or negative pixel is used in
each class in flood vulnerability mapping classification. As those positive (flood) and nega-
tive (non-flood) pixels, which are correctly broken into positive or negative classes, the
actual positives and true negatives are specified. True positives and false negatives, on the
other hand, refer to those positive and negative points which have been erroneously split
(Schumann et al. 2014).
In a ROC graph under the ROC curve (AUC) value, the region clarified the accuracy
of the two groups. For the validation of the five flood susceptible models, this method-
ology was used. Therefore, the ROC curve was plotted by displaying the 1-specificity on
X-axis against the Y-axis sensitivity. The sensitivity means the number of the pixels that
are exactly separated as flood pixels. In the present analysis, both the progress rates and
the forecast rates were reached through the study region preparation and testing datasets
containing several relevant indices of flood susceptibility. In general, AUCs ranging from
0.5 to 0.6 provide an incompetent model when categorizing infinite results. The AUC
value of 0.6 to 0.7, on the other hand, suggests a low output model. In addition to this, a
model with a value of 0.7 to 0.8 AUC indicates moderate efficiency. If the AUC value is
greater than 0.8, the dataset suggests a strong fitness model (Hosseini et al. 2020).

4. Results
4.1. Importance of flood conditioning parameters
Using the VIF values and resistance for each parameter, the estimation of the influence of
flood conditioning variables was evaluated. Among other models, such as Pearson’s cor-
relation coefficients, variance decomposition proportions, conditional index proportions,
VIF and tolerance are the most frequently used models (Belsey 1991). A greater than 10
VIF values or tolerance < 0.1 indicates a collinearity problem with the VIF and tolerance
system analysis. The findings showed that TWI has the highest VIF rating (5.529) and the
lowest tolerance number (0.181) respectively (Table 7). All values are very far from critical
values, so the influencing factors do not have any multicollinearity and can all be used for
modeling. The results of multicollinearity diagnostic tests indicate that TWI with the
highest VIF (5.529) followed by slope, elevation, flow accumulation, LULC, SPI, distance
from the road, distance from the river, rainfall, TRI, aspect, curvature and flow direction
18
I. A. AHMED ET AL.

Table 7. Evaluation of the flood influencing factors using a multicollinearity test (VIF and Tolerance values).
Parameters Elevation Slope Aspect Curvature Flow accumulation Flow direction Distance from river Distance from road Rainfall LULC TRI TWI SPI
VIF 2.87 4.62 1.18 1.08 2.78 1.03 1.30 1.33 1.25 1.94 1.22 5.52 1.72
Tolerance 0.34 0.21 0.84 0.91 0.35 0.96 0.76 0.75 0.79 0.51 0.81 0.18 0.58
GEOCARTO INTERNATIONAL 19

respectively are the crucial potential factors for the frequency of floods in the study area.
The tolerance values of all influencing factors are also more than 0.1, indicating no prob-
lem with multicollinearity.

4.2. Analysis of flood susceptibility mapping


In this study, a machine learning algorithm is combined with four metaheuristic opti-
mization models, i.e. random forest (RF), with ACO, bee colony optimization (BCO),
grey wolf optimization (GWO) and PSO. Determination of the effect of flood influencing
parameters using VIF and tolerance values were developed and used to forecast areas of
susceptibility to flooding in the urban watershed of Guwahati. Flood susceptibility zones
have been categorized into a total of five distinct groups, namely,’ very low’, low,
‘moderate’, high and ’very high’ (Figure 4a–d). The RF-PSO model projected an area of
69.03 km2 as the maximum area of the overall watershed as a ’very highly’ flood suscep-
tible area compared with the smallest area covered by ’moderate’ susceptibility area
(60.87 km2; Table 8). 64.92 km2 and 66.28 km2 areas were predicted by the RF-ACO
model as ’very high’ and ’high’ flood susceptible zones (Figure 4a). In comparison, the

Table 8. Area covered by the four flood susceptible models in the urban watershed of Guwahati.
Area
Very low Low Moderate High Very high
Methods (in km2) (in km2) (in km2) (in km2) (in km2)
RF-ANT 71.60 71.28 65.03 66.28 64.92
RF-BEE 62.80 74.28 65.09 69.36 60.36
RF-WOLF 66.93 70.71 67.03 61.78 65.60
RF-PSO 77.60 67.28 60.87 64.40 69.03

Figure 4. Flood susceptibility models using (a) RF-ant colony (b) RF-bee search (c) RF-grey wolf and (d) RF-par-
ticle swarm.
20 I. A. AHMED ET AL.

bee colony with random forest algorithms predicted an area of 60.36 km2 as a ’very high’
flood susceptible zone. The very low susceptibility to floods was mainly covered by RF-
BCO and the RF-GWO with an area of 62.80 km2 and 66.93 km2, respectively
(Figure 4b,c).
It can be seen in all the maps that there are very high and high flood susceptible zones
in the areas lying in the flat plains of the study area along with proximity to the main
river. In the works of Sarmah et al, (2020) similar results were observed. Also, the ’very
high’ and ’high’ susceptible classes were primarily located in the eastern, southeastern and
middle sections of the urban watershed. A common spatial distribution pattern of flood
susceptible areas has been observed in the case of all the four models in terms of area
under the ’high’ and ’very high’ class. In the case of RF-ACO, though, the figures for
’high’ and ’very high’ classes (38.7%) are comparatively a little lower than the other mod-
els. This suggests a more realistic result, with better precision and reliability, linked to the
bee colony optimization model than other methods.
Settlement in the study area, indicated by red color (Figure 3j) was found in the same
region where ’very high’ and ’high’ flood susceptible areas were located (Figure 4a–d). It
also shows that the low-lying areas especially the city center have maximum ’high’ and
’moderate’ flood susceptibility zones while the hill-tops were found to fall in the ’low’
and ’very low’ susceptibility zones. Yang et al. (2015) and Talukdar et al. (2020) have
stated that their working area had found very similar spatial trends of flood susceptible
zones in urban areas. Yang et al. (2015) also reported that the world’s 0.18 million
population on an annual basis has faced damage due to flood events, specifically those
regions that are closest to the rivers and are mainly densely populated urban and semi-
urban areas.

4.3. Validation and evaluation of flood susceptibility models


The validation of flood-susceptible maps produced using the models was evaluated using
the training point datasets (the success rate curve for floods) and the testing or validation
point (for anticipated flood event the prediction rate curve). That is why the testing and
training flood points were placed on the flood susceptibility maps. To introduce and build
flood susceptibility maps in this analysis, four individual models (ACO, BCO, PSO and
GWO) were used. It is possible to clarify the accuracy of the models using the area ROC
(AUC). Based on the testing points, the four flood susceptible models were validated
using ROC (Table 9). The higher the AUC, the greater the accuracy of the model’s pre-
dictive analysis. To test the assessment of these models, the AUC and a significant level of
the ROC curve were used. In this study, the values of these four individual models were
statistically significant (significant level, .5). Figure 5 revealed that the best-performed
model was the RF-BCO model (AUC¼ 0.856), followed in the test stage by RF-
WOLF (0.841).
Table 10 also presents the analytical success of the four flood susceptible models based
on training data using various error measures such as root mean square error (RMSE),
mean absolute error (MAE), Spearman’s rho and correlation coefficient. The result specif-
ically showed that, of all other models for the training stage, the ant colony model had
the lowest RMSE ¼ 0.167, lowest MAE ¼ 0.1447 and correlation coefficient¼ 0.9725
(Table 10). While the Spearman’s rho, r ¼ 0.8661 for every model. Thus, implying the cor-
relation between the actual flood data and predicted flood data is very high in the
ACO model.
GEOCARTO INTERNATIONAL 21

Figure 5. The validation of four flood susceptible models using ROC based on training point.

Table 9. Accuracy assessment of four flood susceptible models for testing data using the AUC value of the
ROC curve.
Area under the curve
Asymptotic 95% confidence interval
Test result variable(s) Area Std. error Asymptotic sig. Lower bound Upper bound
RF-ANT 0.831 0.049 .000 0.734 0.927
RF-BEE 0.856 0.044 .000 0.770 0.943
RF-WOLF 0.841 0.048 .000 0.747 0.935
RF-PSO 0.819 0.051 .000 0.718 0.920

Table 10. Accuracy assessment of four flood susceptible models for training data using different error measures.
RF-ANT RF-BEE RF-PSO RF-WOLF
Methods
RMSE 0.1678 0.172 0.1738 0.1762
MAE 0.1447 0.1458 0.151 0.1487
Correlation coefficient 0.9725 0.9684 0.9709 0.9656
Spearman’s rho 0.8661 0.8661 0.8661 0.8661

The outcome of statistical performance in the training stage signifies that the ACO
model performed better than all four models, which, in the case of flood susceptibility
modeling reflects a very high level of satisfaction along with fewer errors. In fact, for this
analysis, random forest optimization models with bee colony and ant colony were found
to be the most dominant metaheuristic optimization algorithms with the highest predic-
tion accuracy of all models performed. This result is significantly settled upon for the
assessment of the flood susceptibility mapping, and the process of modeling urban flood
susceptibility with metaheuristic optimization algorithms and machine learn-
ing algorithms.
22 I. A. AHMED ET AL.

5. Discussion
Recurring flood is one of the most disastrous natural events that happen in Assam (Lal
et al. 2020). In addition, Guwahati’s urban watershed is Assam’s capital city, which is situ-
ated in the Brahmaputra Valley and is one of India’s most flood-prone urban centers as
well as the world’s most vulnerable urban centers (Sarmah and Das 2018). Hence, there is
a need to take an approach towards the prediction and mitigation of floods and other nat-
ural disasters to reduce the loss of life as well as property damage in the hazard suscep-
tible areas (Talukdar et al. 2020). Therefore, it is very important to create flood
susceptibility maps and models, as they are very essential to assess risk. Thus, the present
study is completely dedicated to the utilization of new and robust methods to acquire
very accurate results and high precision, which will help in future flood management
practices and plans (Bui et al. 2020; Wang et al. 2020).
The present study used 13 flood influencing parameters to predict the flood susceptibil-
ity of the urban watershed of Guwahati. These flood influencing factors in any area are
relatively complex with several of their mechanisms still under research. The results
clearly show a distinction of the predictive powers of the influencing factors with floods
in the study area. The highest predictive power was observed by TWI when tested
through VIF and tolerance values. Also, high predictive powers were obtained by slope,
elevation, flow accumulation, LULC, SPI, distance from the road, distance from the river,
rainfall, TRI, aspect, curvature and flow direction, respectively. All the factors used in the
study have quite high predictive power for flood events in the study area; hence, none of
them were eliminated in the outcomes of multicollinearity diagnostics tests.
With the help of these thirteen floods influencing factors, the present paper attempted to
integrate four metaheuristic optimization algorithms viz., ant colony, bee search, grey wolf
and particle swarm with a machine learning algorithm, random forest, to create flood sus-
ceptibility maps for the urban watershed of Guwahati, Assam. For a very long time, the
study area has undergone recurrent flooding; therefore, modeling the flood susceptibility of
this region has become very important. Besides, it is evident in all the models that the areas
with proximity to the main river and situated in the flat plains of the study area have very
high and high flood susceptible zones. Similar observations of flood susceptibility in the
Guwahati urban watershed have also been identified in the works of Sarmah et al. (2020).
All the four models used in the study have similar areas under the ’high’ and ’very high’
class and an identical spatial distribution pattern of flood susceptibility is visible in the
study area. The result of this study will be of great assistance to the regional as well as local
authorities of the urban watershed of Guwahati. The study will provide numerous and
extensive approaches to the policy-makers in the mitigation of flood risks and also help in
the development of appropriate measures to avoid potential damages in the study area.
The four applied metaheuristic optimization models have been widely utilized for
researches in flood susceptibility (Khosravi et al. 2019; Ali et al. 2020), landslide suscepti-
bility (Wu et al. 2020), gully erosion (Band et al. 2020) and forest fire (Moayedi et al.
2020). But only a few studies have applied random forest with these models for the pre-
diction of different natural hazards including flood susceptibility. Therefore, the perform-
ance of the four models with the random forest was evaluated and compared by using the
AUC value of the ROC curve at the testing scale and a series of nonparametric statistics
by RMSE, MAE, correlation coefficient and Spearman’s Rho at training scale. Also, using
the VIF and tolerance value, the influences of the flood conditioning parameters
were evaluated.
The findings specifically showed that the suggested random forest-bee colony optimiza-
tion (RF-BCO) models performed better than the other three models in line with the
GEOCARTO INTERNATIONAL 23

Figure 6.

AUC for testing data sets (0.856) (Yariyan et al. 2020). Even though, the other three mod-
els also performed quite good which can be seen from the AUC value achieved (more
than 0.8) at the testing datasets. While, based on the training datasets, the random forest
ant colony optimization (RF-ACO) model out-performed the other three models with low
RMSE (0.1447) and MAE (0.1458) and a very high correlation coefficient value (0.9725).
Therefore, it can be concluded from the results obtained that the RF-BCO and the RF-
ACO are the best two models that performed better in the study.
Because of the recurring urban flood, post-flood events in urban areas are challenging
to witness and manage. According to news reports, floodwaters inundated the Assembly
Secretariat and the State’s Old MLA Hostel complexes. Also, neck-deep water is reported
in some areas of Senduri Ali, Tarun Nagar and Nabin Nagar and the residents, who have
to deal with a recurring flash flood after every heavy downpour. Moreover, the weaker
sections of society are the ones who suffer the most. They have the fewest resources to
deal with the consequences of urban flooding. Loss of hourly wages, for example, or time
lost due to traffic re-routing may only affect middle-class workers, but these add up to
the catastrophic consequences of a disaster for lower-income communities. Paul et al.
(2019) observed a similar trend in Guwahati and assessed community vulnerabilities con-
sidering climate change. Because of its increasing intensity and recurring pattern annually,
urban flooding is gaining importance in the realm of policy-making. Figure 6a–d was
compiled from various sources to depict Guwahati’s urban watershed post-flood events.
24 I. A. AHMED ET AL.

6. Conclusion
In the present research, a combination of metaheuristic optimization algorithms with
machine learning algorithms were developed to generate flood susceptibility modeling of
the urban watershed of Guwahati, Assam. The maximum predictive power and flexibility
were noticed for the RF-BCO (0.856) model followed by the RF-WOLF model (0.841).
However, based on the training data, multiple error measurements such as RMSE, MAE,
spearman’s rho and correlation coefficient were used to check the performance of all
models. The findings stated that based on the training datasets, the RF-ACO model
(RMSE ¼ 0.1678, MAE ¼ 1447, COR ¼ 0.97, Spearman’s Rho ¼ 0.866) is the most cap-
able model for flood susceptible modeling. Though, all the models performed very well
for the mapping of flood susceptibility. Also, all these four models calculated a total area
of approximately 39% as ’high’ and ’very high’ vulnerable zones to flooding.
Overall, the results of the present research suggest that flooding has caused substantial
damage to structures and lives in the study area. If such a pattern of flood susceptibility
persists in the urban watershed of Guwahati, the region could have much more severe
implications in the future. Therefore, the current study using the integration of the meta-
heuristic optimization algorithm in the urban watershed of the Guwahati basin for gener-
ating robust flood susceptible models, which can provide valuable information and
methods to be implemented for counteracting initiatives for local authorities and other
planners to establish successful mitigation as well as urban flood and land-use prevention
strategies. Furthermore, the result of the study will be of great assistance to the authorities
of the study area along with providing important techniques to the policy-makers in order
to mitigate and develop appropriate measures for flood risks and avoid potential damages.
The present study has some limitations, such as the application of moderate resolution
digital elevation model (SRTM) and Landsat 8 OLI, less number of training and testing
data and application of ensemble models. To achieve highly robust FSM, the mentioned
limitation should be tackled by utilizing high resolution DEM and satellite image and
deep learning models, like convolutional neural network.

Acknowledgements
The authors are indebted to the Survey of India for providing the Toposheet from which the city maps
were obtained, and Indian Meteorological Department, Regional Meteorological, Guwahati, Assam for
providing climatic data of the study area. Authors are thankful to the learned reviewers for their scholarly
comments and critical insight which have helped to improve the manuscript many folds.

Disclosure statement
The authors declare that there is no conflict of interest among the authors on any issues.

Funding
The first author is thankful to the University Grants Commission for providing Maulana Azad National
Fellowship (MANF) for the doctoral research.

ORCID
Swapan Talukdar http://orcid.org/0000-0001-6680-9791
Shahfahad http://orcid.org/0000-0001-5868-1062
GEOCARTO INTERNATIONAL 25

Mohd. Rihan http://orcid.org/0000-0002-2533-8433


Atiqur Rahman http://orcid.org/0000-0002-9001-5059

Data availability statement


All the datasets generated during this research work are available from the corresponding author (AR) on
reasonable request.

References
Ahmadlou M, Karimi M, Alizadeh S, Shirzadi A, Parvinnejhad D, Shahabi H, Panahi M. 2019. Flood sus-
ceptibility assessment using integration of adaptive networkbased fuzzy inference system (ANFIS) and
biogeography-based optimization (BBO) and BAT algorithms (BA). Geocarto Int. 34(11):1252–1272.
Ahmed N, Hoque MA, Howlader N, Pradhan B. 2021. Flood risk assessment: role of mitigation capacity
in spatial flood risk mapping. Geocarto Int. 1–21. doi:10.1080/10106049.2021.2002422
Al-Abadi AM. 2018. Mapping flood susceptibility in an arid region of southern Iraq using ensemble
machine learning classifiers: a comparative study. Arab J Geosci. 11(9):1–19.
Alfieri L, Bisselink B, Dottori F, Naumann G, de Roo A, Salamon P, Wyser K, Feyen L. 2017. Global pro-
jections of river flood risk in a warmer world. Earth’s Future. 5(2):171–182.
Ali SA, Parvin F, Pham QB, Vojtek M, Vojtekova J, Costache R, Linh NTT, Nguyen HQ, Ahmad A,
Ghorbani MA. 2020. GIS-based comparative assessment of flood susceptibility mapping using hybrid
multi-criteria decision-making approach, naïve Bayes tree, bivariate statistics and logistic regression: a
case of Topl'a basin, Slovakia. Ecol Indic. 117:106620.
Arabameri A, Saha S, Chen W, Roy J, Pradhan B, Bui DT. 2020. Flash flood susceptibility modelling using
functional tree and hybrid ensemble techniques. J Hydrol. 587:125007.
Arora A, Arabameri A, Pandey M, Siddiqui MA, Shukla UK, Bui DT, Mishra VN, Bhardwaj A. 2021.
Optimization of state-of-the-art fuzzy-metaheuristic ANFIS-based machine learning models for flood
susceptibility prediction mapping in the Middle Ganga Plain, India. Sci Total Environ. 750:141565.
Baghel A. 2016. Causes of urban floods in India: study of Mumbai in 2006 and Chennai in 2015.
International Conference on Disaster and Risk Management: AGORA. At: Sohna, Haryana; p.1-26.
Bahrami Y, Hassani H, Maghsoudi A. 2021. Landslide susceptibility mapping using AHP and fuzzy meth-
ods in the Gilan province, Iran. GeoJournal. 86(4):1797–1816.
Band SS, Janizadeh S, Chandra Pal S, Saha A, Chakrabortty R, Shokri M, Mosavi A. 2020. Novel ensemble
approach of deep learning neural network (DLNN) model and particle swarm optimization (PSO) algo-
rithm for prediction of gully erosion susceptibility. Sensors. 20(19):5609.
Belsey DA. 1991. Conditional diagnostics: collinearity and weak data in regression. New York: Wiley.
Beven KJ, Kirkby MJ. 1979. A physically based, variable contributing area model of basin hydrology.
Hydrol Sci Bull. 24(1):43–69.
Breiman L. 1984. Classification and regression trees. Belmont, CA: Wadsworth International Group.
Breiman L. 2001. Random forests. Mach Learn. 45(1):5–32.

Bui DT, Hoang N-D, Martınez-Alvarez F, Ngo P-TT, Hoa PV, Pham TD, Samui P, Costache R. 2020. A
novel deep learning neural network approach for predicting flash flood susceptibility: a case study at a
high frequency tropical storm area. Sci Total Environ. 701:134413.
Bui DT, Panahi M, Shahabi H, Singh VP, Shirzadi A, Chapi K, Khosravi K, Chen W, Panahi S, Li S, et al.
2018. Novel hybrid evolutionary algorithms for spatial prediction of floods. Sci Rep. 8(1):15364.
Bui DT, Pradhan B, Nampak H, Bui QT, Tran QA, Nguyen QP. 2016. Hybrid artificial intelligence
approach based on neural fuzzy inference model and metaheuristic optimization for flood susceptibi-
litgy modeling in a high-frequency tropical cyclone area using GIS. J Hydrol. 540:317–330.
Chang H-S, Chen T-L. 2016. Spatial heterogeneity of local flood vulnerability indicators within flood-
prone areas in Taiwan. Environ Earth Sci. 75(23):1–14.
Chen W, Hong H, Li S, Shahabi H, Wang Y, Wang X, Ahmad BB. 2019. Flood susceptibility modelling
using novel hybrid approach of reduced-error pruning trees with bagging and random subspace
ensembles. J Hydrol. 575:864–873.
Chen W, Li Y, Xue W, Shahabi H, Li S, Hong H, Wang X, Bian H, Zhang S, Pradhan B, et al. 2020.
Modeling flood susceptibility using data-driven approaches of naïve Bayes tree, alternating decision
tree, and random forest methods. Sci Total Environ. 701:134979.
26 I. A. AHMED ET AL.

Cheng M-Y, Hoang N-D. 2015. Typhoon-induced slope collapse assessment using a novel bee colony
optimized support vector classifier. Nat Hazards. 78(3):1961–1978.
Chou X, Gambardella LM, Montemanni R. 2021. A tabu search algorithm for the probabilistic orienteer-
ing problem. Comput Oper Res. 126:105107.
Clerc M, Kennedy J. 2002. The particle swarm - explosion, stability, and convergence in a multidimen-
sional complex space. IEEE Trans Evol Comput. 6(1):58–73.
Costache R, Tien Bui D. 2020. Identification of areas prone to flash-flood phenomena using multiple-cri-
teria decision-making, bivariate statistics, machine learning and their ensembles. Sci Total Environ.
712:136492.
Cutler A, Stevens JR. 2006. Random forests for microarrays. Methods Enzymol. 411:422–432.
Dodangeh E, Choubin B, Eigdir AN, Nabipour N, Panahi M, Shamshirband S, Mosavi A. 2020. Integrated
machine learning methods with resampling algorithms for flood susceptibility prediction. Sci Total
Environ. 705:135983.
Dorigo M, Blum C. 2005. Ant colony optimization theory: a survey. Theor Comput Sci. 344(2–3):
243–278.
Dormann CF, Elith J, Bacher S, Buchmann C, Carl G, Carre G, Marquez JRG, Gruber B, Lafourcade B,
Leit~ao PJ, et al. 2013. Collinearity: a review of methods to deal with it and a simulation study evaluat-
ing their performance. Ecography. 36(1):27–46.
Duque EL, Aquino PT. 2019. Anthropometric analysis in automotive manual transmission gearshift qual-
ity perception. CTI Symp. 2018:97–109.
El-Haddad BA, Youssef AM, Pourghasemi HR, Pradhan B, El-Shater AH, El-Khashab MH. 2021. Flood
susceptibility prediction using four machine learning techniques and comparison of their performance
at Wadi Qena Basin. Nat Hazards. 105(1):83–114.
El-Magd SA. 2022. Random forest and naïve Bayes approaches as tools for flash flood hazard susceptibil-
ity prediction, South Ras El-Zait, Gulf of Suez Coast, Egypt. Arab J Geosci. 15(3):1–12.
El-Magd SAA, Pradhan B, Alamri A. 2021. Machine learning algorithm for flash flood prediction mapping
in Wadi El-Laqeita and surroundings, Central Eastern Desert, Egypt. Arab J Geosci. 14(4):1–14.
Falah F, Rahmati O, Rostami M, Ahmadisharaf E, Daliakopoulos IN, Pourghasemi HR. 2019. Artificial
neural networks for flood susceptibility mapping in data-scarce urban areas. In: Pourghasemi HR,
Gokceoglu C, editors. ScienceDirect. Elsevier; p. 323–336.
Farfan JF, Cea L. 2021. Coupling artificial neural networks with the artificial bee colony algorithm for glo-
bal calibration of hydrological models. Neural Comput Appl. 33(14):8479–8416.
Hong H, Panahi M, Shirzadi A, Ma T, Liu J, Zhu A-X, Chen W, Kougias I, Kazakis N. 2018. Flood sus-
ceptibility assessment in Hengfeng area coupling adaptive neuro-fuzzy inference system with genetic
algorithm and differential evolution. Sci Total Environ. 621:1124–1141.
Haghighi AT, Darabi H, Karimidastenaei Z, Davudirad AA, Rouzbeh S, Rahmati O, Sajedi-Hosseini F,
Kl€ove B. 2021. Land degradation risk mapping using topographic, human-induced, and geoenviron-
mental variables and machine learning algorithms, for the Pole-Doab watershed, Iran. Environ Earth
Sci. 80(1):1–21.
Hosseini FS, Choubin B, Mosavi A, Nabipour N, Shamshirband S, Darabi H, Haghighi AT. 2020. Flash-
flood hazard assessment using ensembles and Bayesian-based machine learning models: application of
the simulated annealing feature selection method. Sci Total Environ. 711:135161.
Il’inskii NB, Yakimov ND. 1987. Determination of the shape of the downstream slope of an earth-fill dam
from the seepage strength conditions at the face. Fluid Dyn. 22(3):414–419.
Imamura Y. 2022. Development of a method for assessing country-based flood risk at the global scale. Int
J Disaster Risk Sci. 13(1):87–99.
Janizadeh S, Avand M, Jaafari A, Phong TV, Bayat M, Ahmadisharaf E, Prakash I, Pham BT, Lee S. 2019.
Prediction success of machine learning methods for flash flood susceptibility mapping in the Tafresh
watershed, Iran. Sustainability. 11(19):5426.
Javidan N, Kavian A, Pourghasemi HR, Conoscenti C, Jafarian Z. 2020. Data mining technique (max-
imum entropy model) for mapping gully erosion susceptibility in the gorganrood watershed, Iran. In:
Gully Erosion Studies from India and Surrounding Regions. Cham: Springer; p. 427–448.
Jenson SK, Domingue JO. 1988. Extracting topographic structure from digital elevation data for geo-
graphic information-system analysis. Photogramm Eng Remote Sens. 54(11):15931600.
Kalantar B, Ueda N, Saeidi V, Janizadeh S, Shabani F, Ahmadi K, Shabani F. 2021. Deep neural network
utilizing remote sensing datasets for flood hazard susceptibility mapping in Brisbane, Australia. Remote
Sens. 13(13):2638.
Karaboga D, Basturk B. 2008. On the performance of artificial bee colony (ABC) algorithm. Appl Soft
Comput. 8(1):687–697.
GEOCARTO INTERNATIONAL 27

Kashyap S, Mahanta R. 2018. Vulnerability aspects of urban flooding: a review. Indian J Econ Develop.
14(3):578.
Kazakis N, Kougias I, Patsialis T. 2015. Assessment of flood hazard areas at a regional scale using an
index-based approach and Analytical Hierarchy Process: application in Rhodope-Evros region, Greece.
Sci Total Environ. 538:555–563.
Kennedy J, Eberhart RC. 1997. A discrete binary version of the particle swarm algorithm. IEEE Xplore. 5:
4104–4108.
Khan TA, Alam M, Kadir K, Shahid Z, Mazliham MS. 2020. Prior determination of flash floods: artificial
intelligence based predictive analysis using modified cuckoo search. J Comput Theor Nanosci. 17(2):
990–995.
Khosravi K, Nohani E, Maroufinia E, Pourghasemi HR. 2016. A GIS-based flood susceptibility assessment
and its mapping in Iran: a comparison between frequency ratio and weights-of-evidence bivariate stat-
istical models with multi-criteria decision-making technique. Nat Hazards. 83(2):947–987.
Khosravi K, Pham BT, Chapi K, Shirzadi A, Shahabi H, Revhaug I, Prakash I, Tien Bui D. 2018. A com-
parative assessment of decision trees algorithms for flash flood susceptibility modeling at Haraz water-
shed, northern Iran. Sci Total Environ. 627:744–755.
Khosravi K, Shahabi H, Pham BT, Adamowski J, Shirzadi A, Pradhan B, Dou J, Ly H-B, Gr of G, Ho HL,
et al. 2019. A comparative assessment of flood susceptibility modeling using multi-criteria decision-
making analysis and machine learning methods. J Hydrol. 573:311–323.
Lal P, Prakash A, Kumar A. 2020. Google Earth Engine for concurrent flood monitoring in the lower
basin of Indo-Gangetic-Brahmaputra plains. Nat Hazards (Dordr). 104(2):1947–1952.
Lucas AML, Kibler KM. 2016. Integrated Flood Management in developing countries: balancing flood
risk, sustainable livelihoods, and ecosystem services. Int J River Basin Manag. 14(1):19–31.
Meles MB, Younger SE, Jackson CR, Du E, Drover D. 2020. Wetness index based on landscape position
and topography (WILT): modifying TWI to reflect landscape position. J Environ Manage. 255:109863.
Moayedi H, Mehrabi M, Bui DT, Pradhan B, Foong LK. 2020. Fuzzy-metaheuristic ensembles for spatial
assessment of forest fire susceptibility. J Environ Manage. 260:109867.
Moghaddam DD, Pourghasemi HR, Rahmati O. 2019. Assessment of the contribution of geo-environmen-
tal factors to flood inundation in a semi-arid region of SW Iran: comparison of different advanced
modeling approaches. In: Advances in natural and technological hazards research. Vol. 48.
Netherlands: Springer; p. 59–78.
Moore ID, Grayson RB, Ladson AR. 1991. Digital terrain modelling: a review of hydrological, geomorpho-
logical, and biological applications. Hydrol Process. 5(1):3–30.
National Disaster Management Authority of India. 2008. Annual Report 2008. Government of India.
https://ndma.gov.in/sites/default/files/PDF/Reports/ANNUAL%20REPORT%202008-09-ENGLISH.pdf.
Nguyen TS, Nguyen TTT, Bui XT, Chau TD. 2021. Remote sensing and GIS for urbanization and flood
risk assessment in Phnom Penh, Cambodia. Geocarto Int. 1–15. doi:10.1080/10106049.2021.1941307
Niyomubyeyi O, Pilesj€o P, Mansourian A. 2019. Evacuation planning optimization based on a multi-
objective artificial bee colony algorithm. ISP RS Int J Geo-Inform. 8(3):110.
Nhu V-H, Thi Ngo P-T, Pham TD, Dou J, Song X, Hoang N-D, Tran DA, Cao DP, Aydilek IB, _ Amiri M,
et al. 2020. A new hybrid firefly–PSO optimized random subspace tree intelligence for torrential rain-
fall-induced flash flood susceptible mapping. Remote Sens. 12(17):2688.
NIDM. 2011. India Disaster Report 2011. National Institute of Disaster Management. https://nidm.gov.in/
PDF/pubs/India%20Disaster%20Report%202011.pdf.
Nkwunonwo UC, Whitworth M, Baily B. 2020. A review of the current status of flood modelling for
urban flood risk management in the developing countries. Sci Afr. 7:e00269.
Panahi M, Dodangeh E, Rezaie F, Khosravi K, Van Le H, Lee MJ, Lee S, Thai Pham B. 2021. Flood spatial
prediction modeling using a hybrid of meta-optimization and support vector regression modeling.
CATENA. 199:105114.
Panahi M, Gayen A, Pourghasemi HR, Rezaie F, Lee S. 2020. Spatial prediction of landslide susceptibility
using hybrid support vector regression (SVR) and the adaptive neuro-fuzzy inference system (ANFIS)
with various metaheuristic algorithms. Sci Total Environ. 741:139937.
Pandey M, Arora A, Arabameri A, Costache R, Kumar N, Mishra VN, Nguyen H, Mishra J, Siddiqui MA,
Ray Y, et al. 2021. Flood susceptibility modeling in a subtropical humid low-relief alluvial plain envir-
onment: application of novel ensemble machine learning approach. In: D. Ouzounov, Liu JY, Taylor
PT, Hattori K, editors. Front Earth Science. Vol. 9. p. 659296.
Paul A, Deka J, Gujre N, Rangan L, Mitra S. 2019. Does nature of livelihood regulate the urban
community’s vulnerability to climate change? Guwahati city, a case study from North East India. J
Environ Manage. 251:109591.
28 I. A. AHMED ET AL.

Pawe CK, Saikia A. 2018. Unplanned urban growth: land use/land cover change in the Guwahati
Metropolitan Area, India. Geogr Tidsskr. 118(1):88–100.
Pawe CK, Saikia A. 2020. Decumbent development: urban sprawl in the Guwahati Metropolitan Area,
India. Singapore J Trop Geogr. 41(2):226–247.
Pham BT, Luu C, Phong TV, Nguyen HD, Le HV, Tran TQ, Ta HT, Prakash I. 2021. Flood risk assess-
ment using hybrid artificial intelligence models integrated with multi-criteria decision analysis in
Quang Nam Province, Vietnam. J Hydrol. 592:125815.
Pham BT, Pradhan B, Tien Bui D, Prakash I, Dholakia MB. 2016. A comparative study of different
machine learning methods for landslide susceptibility assessment: a case study of Uttarakhand area
(India). Environ Model Softw. 84:240–250.
Pham QB, Chandra Pal S, Saha A, Chowdhuri I, Albanai JA, Janizadeh S, Ahmadi K, Khedher KM, Anh
DT, Duan W. 2021. Current and future projections of flood risk dynamics under seasonal precipitation
regimes in the Hyrcanian Forest region. Geocarto Int. 1–18.
Rubinato M, Nichols A, Peng Y, Zhang J, Lashford C, Cai Y, Lin P, Tait S. 2019. Urban and river flood-
ing: comparison of flood risk management approaches in the UK and China and an assessment of
future knowledge needs. Water Sci Eng. 12(4):274–283.
Sahoo SN, Sreeja P. 2017. Development of flood inundation maps and quantification of flood risk in an
urban catchment of Brahmaputra river. ASCE-ASME J Risk Uncertain Eng Syst A: Civil Eng. 3(1):
1–11.
Sarmah T, Das S. 2018. Urban flood mitigation planning for Guwahati: a case of Bharalu basin. J Environ
Manage. 206:1155–1165.
Sarmah T, Das S, Narendr A, Aithal BH. 2020. Assessing human vulnerability to urban flood hazard using
the analytic hierarchy process and geographic information system. Int J Disaster Risk Reduct. 50:
101659.
Schumann GJ-P, Vernieuwe H, De Baets B, Verhoest NEC. 2014. ROC-based calibration of flood inunda-
tion models. Hydrol Process. 28(22):5495–5502.
Shafapour Tehrany M, Kumar L, Neamah Jebur M, Shabani F. 2019. Evaluating the application of the
statistical index method in flood susceptibility mapping and its comparison with frequency ratio and
logistic regression methods. Geomatics Nat Hazards Risk. 10(1):79–101.
Shafizadeh-Moghadam H, Valavi R, Shahabi H, Chapi K, Shirzadi A. 2018. Novel forecasting approaches
using combination of machine learning and statistical models for flood susceptibility mapping. J
Environ Manage. 217:1–11.
Shahabi H, Shirzadi A, Ronoud S, Asadi S, Pham BT, Mansouripour F, Geertsema M, Clague JJ, Bui DT.
2021. Flash flood susceptibility mapping using a novel deep learning model based on deep belief net-
work, back propagation and genetic algorithm. Geosci Front. 12(3):101100.
Sun D, Wen H, Wang D, Xu J. 2020. A random forest model of landslide susceptibility mapping based
on hyperparameter optimization using Bayes algorithm. Geomorphology. 362:107201.
Talukdar S, Ghose B, Shahfahad Salam R, Mahato S, Pham QB, Linh NTT, Costache R, Avand M. 2020.
Flood susceptibility modeling in Teesta River basin, Bangladesh using novel ensembles of bagging algo-
rithms. Stochastic Environ Res Risk Assess. 34(12):2277–2300.
Tehrany MS, Jones S, Shabani F. 2019. Identifying the essential flood conditioning factors for flood prone
area mapping using machine learning techniques. CATENA. 175:174–192.
Tehrany MS, Pradhan B, Jebur MN. 2014. Flood susceptibility mapping using a novel ensemble weights-
of-evidence and support vector machine models in GIS. J Hydrol. 512:332–343.
Termeh SVR, Kornejady A, Pourghasemi HR, Keesstra S. 2018. Flood susceptibility mapping using novel
ensembles of adaptive neuro fuzzy inference system and metaheuristic algorithms. Sci Total Environ.
615:438–451.
Termeh SVR, Shirani K, Pasandi M. 2021. Mapping of landslide susceptibility using the combination of
neuro-fuzzy inference system (ANFIS), ant colony (ANFIS-ACOR), and differential evolution (ANFIS-
DE) models. Bull Eng Geol Environ. 80(3):2045–2067.
Tien BD, Tuan TA, Hoang ND, Thanh NQ, Nguyen DB, Van Liem N, Pradhan B. 2017. Spatial predic-
tion of rainfallinduced landslides for the Lao Cai area (Vietnam) using a hybrid intelligent approach of
least squares support vector machines inference model and artificial bee colony optimization.
Landslides, 14(2):447–458.
Tripathi G, Pandey AC, Parida BR. 2022. Flood hazard and risk zonation in north Bihar using satellite-
derived historical flood events and socio-economic data. Sustainability. 14(3):1472.
Wang Y, Fang Z, Hong H, Peng L. 2020. Flood susceptibility mapping using convolutional neural net-
work frameworks. J Hydrol. 582:124482.
GEOCARTO INTERNATIONAL 29

Wang Y, Hong H, Chen W, Li S, Panahi M, Khosravi K, Shirzadi A, Shahabi H, Panahi S, Costache R.


2019. Flood susceptibility mapping in Dingnan County (China) using adaptive neuro-fuzzy inference
system with biogeography based optimization and imperialistic competitive algorithm. J Environ
Manage. 247:712–729.
Wu Y, Ke Y, Chen Z, Liang S, Zhao H, Hong H. 2020. Application of alternating decision tree with
AdaBoost and bagging ensembles for landslide susceptibility mapping. CATENA. 187:104396.
Yadav R, Barua A. 2018. Contribution of urbanization to emissions: case of Guwahati City, India. Water
Sci Technol Lib. 84:69–96.
Yang YCE, Ray PA, Brown CM, Khalil AF, Yu WH. 2015. Estimation of flood damage functions for river
basin planning: a case study in Bangladesh. Nat Hazards. 75(3):2773–2791.
Yariyan P, Janizadeh S, Van Phong T, Nguyen HD, Costache R, Van Le H, Pham BT, Pradhan B,
Tiefenbacher JP. 2020. Improvement of best first decision trees using bagging and dagging ensembles
for flood probability mapping. Water Resour Manage. 34(9):3037–3053.
Yin J, He F, Xiong YJ, Qiu GY. 2017. Effects of land use/land cover and climate changes on surface run-
off in a semi-humid and semi-arid transition zone in northwest China. Hydrol Earth Syst Sci. 21(1):
183–196.
Yu C, Chen J. 2020. Landslide susceptibility mapping using the slope unit for southeastern Helong City,
Jilin Province, China: a comparison of ANN and SVM. Symmetry. 12(6):1047.
Zhang H, Yao Z, Yang Q, Li S, Baartman JEM, Gai L, Yao M, Yang X, Ritsema CJ, Geissen V. 2017. An
integrated algorithm to evaluate flow direction and flow accumulation in flat regions of hydrologically
corrected DEMs. CATENA. 151:174–181.
Zhang X, Cao W, Guo Q, Wu S. 2010. Effects of landuse change on surface runoff and sediment yield at
different watershed scales on the Loess Plateau. Int J Sediment Res. 25(3):283–293.
Zhang P, Wu HN, Chen RP, Chan TH. 2020. Hybrid meta-heuristic and machine learning algorithms for
tunneling-induced settlement prediction: a comparative study. Tunnell Underg Space Technol. 99:
103383.

You might also like