You are on page 1of 14

Journal of Hydrology 591 (2020) 125552

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Research papers

Convolutional neural network approach for spatial prediction of flood T


hazard at national scale of Iran
Khabat Khosravia, Mahdi Panahib,c, Ali Golkariand, Saskia D. Keesstrae,f, Patricia M. Sacof,
Dieu Tien Buig, , Saro Leeb,h,
⁎ ⁎

a
Department of Watershed Management Engineering, Sari Agricultural Science and Natural Resources University, Sari, Iran
b
Geoscience Platform Division, Korea Institute of Geoscience and Mineral Resources (KIGAM), 124, Gwahak-ro Yuseong-gu, Daejeon 34132, South Korea
c
Division of Science Education, Kangwon National University, College of Education, # 4-301, Gangwondaehak-gil, Chuncheon-si, Gangwon do 24341, South Korea
d
Department of Watershed Management, Ferdowsi University of Mashhad, Mashhad, Iran
e
Team Soil, Water and Land Use, Wageningen Environmental Research, Wageningen University & Research, Wageningen, Netherlands
f
Civil, Surveying and Environmental Engineering and Centre for Water Security and Environmental Sustainability, The University of Newcastle, Callaghan 2308, Australia
g
Institute of Research and Development, Duy Tan University, Da Nang 550000, Viet Nam
h
Department of Geophysical Exploration, Korea University of Science and Technology, 217 Gajeong-ro Yuseong-gu, Daejeon 34113, South Korea

ARTICLE INFO ABSTRACT

This manuscript was handled by Emmanouil Iran experiences frequent destructive floods with significant socioeconomic consequences. Quantifying the ac­
Anagnostou, Editor-in-Chief curate impacts of such natural hazards, however, is a complicated task. The present study uses a deep learning
convolutional neural networks (CNN) algorithm, which is among the newer and most powerful algorithms in big
Keywords:
Flood control data sets, to develop a flood susceptibility map for Iran. A total of 2769 records were collected from flood
Hazard locations across the entire country; we divided this data set into two groups using a cross-validation technique.
Iran The first group, used as a training data set, was constructed from 70% of the data set and was used for model
Deep-learning building. The second group, used as a testing data set, was constructed from the remaining 30% of the records
and used for validation. Ten flood conditioning factors, slope, altitude, aspect, plan curvature, profile curvature,
rainfall, geology, land use, distance from roads, and distance from rivers, were identified and used in the
modeling process. The area under the prediction-rate curve was used for model evaluation, with results showing
that the flood susceptibility map has an acceptable accuracy of 75%. The results also indicated that approxi­
mately 12% and 3% of the country are highly and very highly susceptible to future flooding events, respectively.
Moreover, 29% and 49% of Iran’s cities are located in areas with high and very high susceptibility to future
flooding hazards. The most effective approaches to flood mitigation are preventing urban expansion and new
construction in highly to very highly flood-prone areas as well as watershed management plans and constructing
flood control structures according to the topographical characteristics of the catchment.

1. Introduction 2016a, 2016b; Termeh et al., 2018; Tiryaki and Karaca, 2018).
Floods affect 75 million people on average annually worldwide
Floods are among the most destructive natural hazards in tropical (around 20,000 casualties per year), with economic losses of $95 billion
areas that experience high rainfall, in particular during monsoon per­ (Chapi et al., 2017; Khosravi et al., 2018b). Recently, the frequency of
iods, and in dry and semi-dry regions (as flash floods). Flooding can flooding has increased by more than 40% (Hirabayashi et al., 2013)
cause irreparable and immeasurable damage that may be categorized as because climate change has significantly influenced the intensity, pat­
tangible or intangible. Tangible damage, which includes destruction to tern, and magnitude of flooding (Chapi et al., 2017; Hens et al., 2018;
buildings, agriculture, road and transportation systems, infrastructure, Van Thanh et al., 2017). Asian countries, in particular, are hotspots,
environmental ecosystems, humans, and property, imposes a major with relatively high exposure to flooding and earthquakes. In Asia,
negative emotional burden on the society in question (Khosravi et al., more than 90% of human losses resulting from natural disasters are

Corresponding authors at: Geoscience Platform Division, Korea Institute of Geoscience and Mineral Resources (KIGAM), 124, Gwahak-ro Yuseong-gu, Daejeon

34132, South Korea (S. Lee).


E-mail addresses: mahdi.panahi@kigam.re.kr (M. Panahi), golkarian@um.ac.ir (A. Golkarian), saskia.keesstra@wur.nl (S.D. Keesstra),
Patricia.Saco@newcastle.edu.au (P.M. Saco), buitiendieu@duytan.edu.vn (D.T. Bui), leesaro@kigam.re.kr (S. Lee).

https://doi.org/10.1016/j.jhydrol.2020.125552
Received 23 March 2020; Received in revised form 11 August 2020; Accepted 15 September 2020
Available online 20 September 2020
0022-1694/ © 2020 Elsevier B.V. All rights reserved.
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

Fig. 1. Photos of flood damage in a different part of Iran in 2019.

attributable to flooding (Smith, 2003). Flooding results from the direct of data that they require for modeling and calibration are unavailable in
impact of both natural and anthropogenic factors (Chang and Chen, many developing countries. These models are also computationally
2016). Iran has recently experienced disastrous floods that, in addition demanding and time-consuming, in particular in large areas (i.e., gen­
to the impact of climate change, can be attributed to poor watershed erating a large-scale flood map, such as for an entire country, is diffi­
management (e.g., deforestation, overgrazing, lack of flood control cult). In general, two- or three-dimensional flood modeling is unviable
measures; (Khosravi et al., 2016a). Recent flooding events in Iran and for areas larger than 1000 km2, in particular, when high resolution is
their record casualties have affected 25 of the country’s 31 provinces, required (Teng et al., 2017). Several powerful physically-based models
resulting in more than 77 human casualties and USD $2.2 billion worth (hydrographs) have been developed to predict flood discharge (e.g.,
of damage to public and private assets (https://en.wikipedia.org/wiki/ SWAT, HEC-HMS). Although these models are robust, well known, and
2019_Iran_floods). Some real-photo of the flooding events in a different developed according to the relevant catchments’ rainfall-runoff pro­
part of the country is presented in Fig. 1. Furthermore, at the time of cesses, they still have several shortcomings for generating flood maps
this writing, a vast region in southern and southeastern Iran (including and identifying inundation areas (Sahoo et al., 2009). Consequently,
Sistan & Baluchistan) is facing unprecedented floods. researchers are keen to identify alternative robust, and flexible models
Although flooding hazards are inevitable and largely beyond human to remedy these weaknesses.
control, flood risk management practices can mitigate the impact of Recently, the use of geographic information systems (GIS) and re­
flooding through hazard prediction and vulnerability reduction. mote sensing (RS) has brought new insights to the field of hydrology
Identifying areas prone to flooding can inform threatened communities and multidimensional phenomena—in particular, flood modeling—to
and decision-makers at different levels to avoid catastrophes and address the weaknesses of the models above. GIS is a useful geostatis­
therefore is key to flood risk management (Bui et al., 2018; Choubin tical tool that can handle large amounts of data of different scales and
et al., 2019; Kourgialas and Karatzas, 2011; Sarhadi et al., 2012; from different sources for (flood) map generation. Several researchers
Tehrany et al., 2013; Tien Bui et al., 2018). Flooding, a complex phe­ have integrated various data-driven (bivariate and multivariate) and
nomenon resulting from watershed management and hydrological, multi-criteria decision-making (MCDM) algorithms with GIS and RS for
meteorological, geomorphological, and geological conditions, is a so­ flood susceptibility mapping (FSM), including frequency ratio (Khosravi
phisticated and non-linear process and thus cannot be easily predicted et al., 2016a; Lee et al., 2012; Tehrany et al., 2013), weights of evidence
with simple linear models (Pappenberger et al., 2006). Numerical (Rahmati et al., 2016), evidential belief function (Tien Bui et al., 2019),
models, such as the Hydrologic Engineering Center’s (CEIWR-HEC) Shannon entropy (Haghizadeh et al., 2017), analytic hierarchy process
River Analysis System (HEC-RAS), offer some advantages, such as the (AHP) (Chen et al., 2011; Das, 2019), analytic network process (ANP)
calculation of discharge in any cross-section, flow depth, or velocity, in (Dano et al., 2019; de Brito and Evers, 2016), the decision-making trial
particular concerning inundation maps (two- or three-dimensional) and and evaluation laboratory (DEMATEL) (Azareh et al., 2019), Technique
are appropriate modeling flood challenges. However, the large amounts for Order Preference by Similarity to Ideal Solution (TOPSIS) (Khosravi

2
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

et al., 2019), and logistic regression (LR) (Youssef et al., 2015). Al­ 30″ E and latitude 25° and 40° N. It is bordered by Iraq and Turkey to
though bivariate statistical models have simple structures and are easy the west and by Armenia, Azerbaijan, the Caspian Sea, and
to interpret, as a result of this simplicity, they are not sufficiently ac­ Turkmenistan to the north. Afghanistan and Pakistan border Iran to the
curate for predicting complex phenomena, such as flood occurrences. east, with the Persian Gulf and the Oman Sea to the south (Fig. 2).
Furthermore, the MCDM technique, as a subjectivity model, is wea­ Almost 50% of Iran consists of mountains; fertile plains account for
kened by its reliance on expert opinion as an essential input, which 25%, and the remaining 25% is desert (Dehnavi et al., 2015). The slope
often introduces bias (Khosravi et al., 2018b, 2019). Thus, machine angle ranges from 0° to 76.28°, and the altitude ranges from 500 to
learning algorithms were developed to address these weaknesses. These 5597 m above sea level. The country’s two main mountains are Zagros
include artificial neural networks (ANN) (Falah et al., 2019; Kia et al., in the west and Alborz in the north. Iran’s climate is unique by virtue of
2012), support vector machines (SVM) (Kavzoglu et al., 2014), random the difference between the warmest and coldest temperatures in winter,
forests (Chapi et al., 2017), logistic model trees (Pham et al., 2018), which can sometimes be more than 50 °C. Iran is located in an arid/
naïve Bayes trees (Khosravi et al., 2019) and multivariate adaptive semi-arid region with annual average precipitation close to 200 mm
regression splines (MARS) (Dodangeh et al., 2020). Each of these al­ and significant spatial variability, from 15 mm/year in the central de­
gorithms has several drawbacks. The use of SVM has declined as a result serts up to 2000 mm/year in the north. The spatial variability in pre­
of the numerous parameters in its structure, for each of which optimum cipitation and temperature has a significant impact on the vegetation
values must be determined, and the need to determine the most effec­ cover. In terms of its geomorphological characteristics, Iran is divided
tive kernel functions in its structure. Although the decision tree (DT) into the different regions of Caspian, Zagros, the central plateau, and
algorithm offers reasonable precision in FSM (Tehrany et al., 2013), it southern coastal plains (Ghorbani, 2013). Political boundaries divide it
has some problems for defining rules for performance. In DT, multiple into 31 provinces and 1243 large cities, 339 of which have popula­
output attributes are not allowed, and trees are sensitive to noisy data tions > 20,000.
(Tien Bui et al., 2012). Another weakness of DT is that this model may
produce a large tree structure, which makes interpretation difficult, and 2. Methodology
misclassification of data may occur during the machine learning process
(Tehrany et al., 2013). ANN, the most widely used algorithm, also A diagram of the conceptual modeling methodology adopted in the
suffers from several drawbacks that result in its poor prediction capa­ current study is presented in Fig. 3.
city. Therefore, ANN was integrated with fuzzy logic, and the adaptive
neuro-fuzzy inference system (ANFIS) was developed. Although the
ANFIS is a powerful model, it struggles to accurately determine the 2.1. Flood inventory mapping
weights of membership functions in fuzzy rules. Some researchers have
integrated different optimization algorithms with the ANFIS model to The first step in the flood modeling and mapping process is to
address this weakness with respect to different aspects of hydrology, compile historical flood data and to determine flood conditioning fac­
such as FSM (Bui et al., 2018, 2016; Khosravi et al., 2018a; Tien Bui tors as well as the spatial relationships between these factors and
et al., 2018; Wang et al., 2019), groundwater potential mapping (Chen flooded locations to understand the dynamics of flooding in the area of
et al., 2019; Khosravi et al., 2018a; Termeh et al., 2018), and landslide interest. A total of 2769 historical flood records collected by Iran’s
susceptibility mapping (Chen et al., 2017), but this successive integra­ Water Resources Management Company (IWRMC) were used for
tion of the ANFIS model with optimization algorithms is complex and modeling. These locations were ultimately converted into several points
has decreased its use. Deep learning algorithms have recently gained and then precisely located in polygonal format.
superiority over traditional machine learning models by virtue of their
greater flexibility and predictive accuracy, in particular for big data, 2.2. Preparation and selection of flood conditioning factors
and incremental approaches have been used to learn high-level com­
ponents from the data sets used (Ghorbanzadeh et al., 2019). Deep In the first stage of modeling, ten flood conditioning factors were
learning algorithms outperform all other traditional machine learning considered based on the literature (Khosravi et al., 2016b; Tehrany
algorithms when a data set is particularly large and the problem under et al., 2015; Wang et al., 2019), catchment characteristics and data
investigation has a complex process (Panahi et al., 2020). A convolu­ availability. These factors were slope, altitude, aspect, plan curvature,
tional neural network (CNN), an automatic tool, is the most popular profile curvature, rainfall, geology, land use, distance from roads, and
deep learning algorithm which is proper for spatial data, like present distance from rivers (Bui et al., 2018, 2016; Khosravi et al., 2016a,
study. This algorithm sets the features as input and can be described as 2018b, 2016b, 2019; Tehrany et al., 2013; Tien Bui et al., 2018, 2019).
a tensor with which local components are inter-correlated. Convolu­ We implemented a widely used approach, the information gain ratio
tional layer in CNN algorithm can simultaneously learn the connection (IGR) technique, to distinguish between variables with a significant
between features and this would result to achieve higher performance. impact on the accuracy of estimation and noise-creating variables with
Also CNN use minimal amount of preprocessing. negative effects on the outcome (Table 2).
In this paper, we propose a framework with a robust deep learning Slope, altitude, aspect, plan curvature, and profile curvature were
algorithm that helps to generate flood susceptibility maps. This fra­ constructed from topographic maps obtained from the National
mework was used to generate maps for Iran and draw conclusions re­ Cartographic Center of Iran (NCC), with a spatial resolution of
garding the spatial distribution of susceptibility and its potential dri­ 30 × 30 m. Distance from rivers and road were computed with data
vers. from Iranian rivers and roads (a shapefile provided by IWRMC) and
using the multiple ring buffer tool in ArcGIS 10.2. The geological map
of Iran was acquired from Geological Survey & Mineral Explorations of
1.1. Study area Iran (GSI). The rainfall map used in this study was constructed by Iran’s
Meteorological Organization (IMO) using long-term rainfall data
Iran is a Middle Eastern country with a population of 82 million and (30 years of mean annual rainfall). Landuse map was obtained (as a
an area of 1,648,195 km2. Iran lies between longitude 44° and 63° 5′ shapefile) from the Iranian Agricultural Scientific Ministry (Table 1).

3
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

Fig. 2. Study area and flood inventory map.

Fig. 3. Flowchart of the methodology adopted in the present study.

Table 1
Input variables description.
Conditioning factor Source Spatial resolution Importance

Aspect DEM 30 m × 30 m Different amount of soil moisture due to different solar radiation and vegetation cover
Slope DEM 30 m × 30 m Infiltration process and runoff generation
Plan and profile curvature DEM 30 m × 30 m Erosion process and control surface runoff
Altitude DEM 30 m × 30 m Has effect on rainfall and vegetation cover as well as soil depth and texture and finally runoff generation
Rainfall National map 30 m × 30 m Affects runoff generation
Distance from roads National map 1:25,000 Cutting the slope hill, river, forest and etc. through hydraulic structures (culverts, bridges). Sometimes is a
manmade cause of flooding.
Distance from rivers National map 1:100,000 Effect on the soil moisture as well as water level increases and extended from rivers to floodplain and cause
flooding.
Geology National map 1:100,000 Different formations have a different structures and permeability.
Land-use National map 1:25,000 Different land-use and vegetation covers affect infiltration rates (and runoff generation).

4
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

Fig. 4. Structure of the deep CNN algorithm.

Table 2 All the input maps and their classes are presented in Fig. A1 in the
Parameter settings for the CNN. Appendix.
Parameter Search space Final settings
2.3. Building of the training and testing data sets
Convolutional kernel size None 10 × 1
Number of convolution unit 100
Max pooling kernel size 3×1 The final step in FSM by artificial intelligence algorithms involves
Number of epochs 90 binary classification, with two classes of flood and non-flood locations.
Activation function ReLU Consequently, ArcGIS10.2 was used to identify 2769 non-flood loca­
Optimizer Adamax tions in places such as areas which are never inundated during flood
Learning rate 0.001
events (one of the most obvious examples is hilly or mountainous areas)
Initial learning rate 0.1
Dropout rate 0.5 (all non-flood locations presented in Fig. B1 in the Appendix). Flood
Weight decay 0.0001 and non-flood locations were assigned values of 1 and 0, respectively,
and were overlaid with the 10 flood conditioning factors. Finally, the
required information (i.e., flood position, located in which class of each
Table 3 flood conditioning factor) was extracted and converted into an Excel
Networks’s accuracy without using a dropout layer to avoid the over-fitting file. All information extracted with respect to flood and non-flood lo­
problem. cations was divided into two groups by a cross-validation technique at a
SSE MSE RMSE MAE r ratio of 70:30, the most commonly used ratio in spatial modeling of
natural hazards (Tehrany et al., 2014, 2015). The first portion was used
Training 34.031 0.009 0.098 0.040 0.982
for training (the training data set) and was constructed from 70% of
Testing 14.941 0.009 0.094 0.040 0.983
flood and non-flood locations for model building, and the second por­
*Sum of squared errors (SSE), mean square error (MSE), root mean square error tion (the testing data set) was used for model validation and was con­
(RMSE), mean absolute error (MAE) and correlation coefficient (r). structed from the remaining 30% of flood and non-flood locations
(Chung and Fabbri, 2003).
Table 4
Networks’s accuracy after using dropout layer to avoid the over-fitting problem. 3. Model description
SSE MSE RMSE MAE r
3.1. Convolutional neural networks (CNN)
Training 28.421 0.008 0.090 0.032 0.984
Testing 14.805 0.008 0.094 0.033 0.098
To overcome the challenge of decreased speed in the learning pro­
cess faced by traditional artificial neural networks (ANN) in analyzing
complex networks (Lu et al., 2017), Bengio proposed the convolutional
neural network (CNN), which is a neural network that carries out local
connections among layers (LeCun et al., 1998; Lu et al., 2017). CNN has
different neuronal layers. The first layer extracts features from an input
image as well as retains connections between pixels by understanding
image features that use small squares of input data. This layer is called
the convolution layer. The convolution layer is a mathematical opera­
tion base and requires two inputs—an image matrix and a filter or
kernel—to be run. The second layer is the activation layer, which
usually appears after the convolutional layer if it is non-linear. How­
ever, a rectified linear unit (ReLU) function is the most effective and
typical activation function and is always zeros and ones. The third
layer, or the pooling layer, preserves the most critical information while
reducing the number of parameters, in particular when images are very
large. Spatial pooling has different types, such as Max, Average, Sum,
and so forth. The last layer is the actual component and performs dis­
criminative learning in a deep neural network. This fully connected
Fig 5. Convergence plot of the models.

5
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

Table 5
Importance of flood conditioning factors using the information gain ratio technique.
Factors Land-use aspect altitude Plan curvature Distance from river rainfall Slope Distance from road Profile curvature geology

IGR value 0.60 0.40 0.37 0.35 0.33 0.30 0.29 0.28 0.27 0.08

Fig. 6. Flood susceptibility map of Iran. The numbers correspond to the different provinces as listed in Fig. 7.

Fig. 7. Percentages of flood-prone areas in each province.

layer is a simple and multilayer perceptron that can learn weights and notable results using CNN-2D, and thus this structure was used in this
identify object classes. Based on the type of data, images, and goals, study (Wang et al., 2019). Moreover, because CNN input data must take
scientists have proposed different structures for their research, in­ the form of images, and the input data are one-dimensional, the pri­
cluding ZFNet, VGGNet, GoogleNet, ResNet, LeNet-5, and AlexNet mary data must be converted into images.
(Fang et al., 2019). Geoscience researchers have achieved and reported To initialize FSM, a one-dimensional input grid cell that consists of

6
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

Fig. 8. The locations of cities (indicated by


dots) considered in this study. The color of
the dot indicates the class of flood suscept­
ibility. The chart on the right shows the
percentage of cities exposed to varying de­
grees of flood hazard. (For interpretation of
the references to colour in this figure legend,
the reader is referred to the web version of
this article.)

Fig. 9. Percentage of flood exposure in each province by city.

Fig. 10. Images showing the construction of a flood wall by locals.

7
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

Fig. 11. Images of flooded cities close to large rivers: (a) Ahvaz in the Khozestan province and (b) Agh-Ghala in the Golestan province.

different attribute features must be converted into a two-dimensional The final approach to avoid over-fitting is by applying L2
matrix. Because none of the data are classified and are constant, the Regularization. In this method, Las function is being added to the
images created are very large. second norm of network weight. This would reduce network overfitting.
The network used has a convolutional layer with 100 filters at the Additionally, the second norm coefficient of weights (weight decay)
beginning, and in most cases, these layers are used for those data that was considered 0.0001. Fig. 5 shows the convergence plot of the
have more correlation with adjacent features. Inherently these kinds of models.
cases are applied for the data-like image. However, it is not useful for
the features used in this article. In order for the network to be able to 3.2. Flood map generation, validation, and hazard assessment
learn the correlation between features simultaneously during the
learning process, the filter size considered should be 10 × 1. This en­ In the present study, a newly developed artificial intelligence deep
ables the filter to consider all features and their correlations, to extract CNN algorithm was used to delineate flood-prone areas. As a first step,
the correlation features. This layer uses the activation function ReLU. the CNN algorithm was trained (model learning) using a training data
The next layer used is the normalization layer that accelerates the set (i.e., the training data set in Excel format used for model building).
learning process. In the following, a dropout layer with 50% rate is Next, the country of Iran was converted into a raster format. The model
used. This layer transforms 50% of the output from pre-layer to 0 values developed was used to predict a flood susceptibility/probability index
in each learning iteration so that the next layer does not overfit learning for each pixel in the study area (i.e., in this stage, the entire country was
inputs, and enable the learning of different aspects of the data. The next used as a test data set for the built model). Finally, these indices were
layer is the average pooling layer with filter size 3 × 1 that reduces its divided into different classes of very low, low, moderate, high, and very
dimension by a half. This layer dramatically reduces the weight of the high susceptibility to flooding. Different methods may be used for this
network and highly decreases its computational load during the classification, but as our predicted indices were skewed, the quantile
learning process. method was used (Khosravi et al., 2019). The final step was to evaluate
The following layer is fully connected, it has five neurons and its the resulting map. Therefore, the testing data set was applied in terms
activation function is a ReLU like convolutional layer. Finally, the last of prediction-rate curves so that the area under the curve (AUC) iden­
layers are also fully connected with one neuron and one regression tified the model prediction capability quantitatively (Khosravi et al.,
layer that produces the final output. The CNN structure and its para­ 2016a, 2016b). Finally, by overlaying the flood susceptibility map on
meters are presented in Fig. 4 and Table 2. the locations of cities in Iran, we identified the number of cities in each
There are several different approaches developed to avoid over-fit­ class and used them for further analyses.
ting, (Srivastava et al., 2014). The first method is to use a set of data
validation (10% of train data) to check the method’s accuracy after 4. Results and discussion
training, which not only adjusts network parameters but also helps to
avoid over-fitting by comparing the network accuracy of data and The flooding prediction map was evaluated with the testing data set.
training data. The second approach is applying a Dropout layer after The results from this phase show a reasonable and good fit, determined
each convolutional layer, which deletes 50% of its input layers after by an area under the prediction-rate curve (AUC) of 0.75. This means
every iteration in order for the network to learn all aspects of the given that the map’s accuracy is around 75% (Khosravi et al., 2019), which is
dataset as well as to avoid over-fitting. The other approach is to use a reliable and acceptable, in particular given the large size of the study
relatively small network which would help the network not to be over- area based on the classification efficiency proposed by Hosmer and
fitted. Lemeshow (2000). Model performance can be classified into different
Tables 3 and 4 show the network’s accuracy for training and testing accuracy classes, including 0.5–0.6, 0.6–0.7, 0.7–0.8, 0.8–0.9, and
data both with and without using a dropout layer. As can be seen, when 0.9–1, which correspond to poor, average, good, very good, and ex­
the dropout layer is not used, network accuracy (over-fitting) of cellent performance, respectively (Yesilnacar, 2005). Given its good
training data is improved. However, for testing data, the difference accuracy, this map was used for further analyses.
between using and not using the dropout layer is quite clear, and it is The results of the information gain ratio method (shown in Table 5)
clear that not using the dropout layer decrease the network accuracy of indicate that the land use factor has the greatest impact on Iran’s flood
testing data, which means this approach is efficient. occurrences (0.60), followed by aspect (0.40), altitude (0.37), plan

8
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

curvature (0.35), distance from rivers (0.33), rainfall (0.30), slope plans and measures are required immediately. These measures can be
(0.29), distance from roads (0.278), and geology (0.08). This demon­ divided into two main groups—structural and nonstructural—or a
strates that all thematic layers considered significantly influence combination of both may be more effective (e.g., structural and biolo­
flooding. By identifying the impact of each factor on flood occurrences, gical). The most important measures are the generation of flood hazard
one can more effectively implement flood control plans and measures. maps, warning systems, floodwalls, dikes, and detention dams; the
Visual inspection of the effective flood conditioning factors (i.e., slope, implementation of land use planning strategies; afforestation; and the
curvature, rainfall, and altitude maps) (Fig. A1) and the flood sus­ planting of appropriate vegetation.
ceptibility map (Fig. 6) clearly reveals that most flood-prone areas are Recent flooding events in Iran in 2019 affected 25 of the country’s
located in regions where mountain ranges are in close proximity to flat 31 provinces, robustly corroborating our findings. The severity of da­
areas, provided these areas have sufficient rainfall. This observation is mage was so great that several countries, including France, Germany,
consistent with the nature of the mechanisms that promote flooding. It Italy, India, South Korea, Turkey, Iraq, Kuwait, Qatar, and Pakistan,
is worth mentioning that although these areas are exposed to high flood among others, sent humanitarian aid to Iran. Although it has been ar­
hazards, they tend to be densely populated because of their fertile soils, gued that most of this damage resulted from the sudden and severe
good weather conditions, and surface water supply, among other fac­ rainfall (https://en.wikipedia.org/wiki/2019_Iran_floods), other stu­
tors. Land use is the factor that most influences flood occurrences be­ dies suggest that poor management catchment strategies and defor­
cause most of Iran’s land is covered by bare soil that is completely estation (Raziei and Zand, 2019), lack of budget for implementing
compacted because of overgrazing and therefore is very effective at watershed management plans, and poor crisis management are the
generating runoff. Land use effectiveness, especially plant's role in most important factors in Iran’s frequent flood occurrences and its da­
runoff generation, is described by (Hou et al., 2020). Although flooding mages. Also Yadollahie (2019) declared that integration of aforemen­
is linked to rainfall, rainfall is not the main factor predicting flood tioned reasons along with global warming is the main cause of recent
susceptibility in Iran, as it is very high in mountainous areas, where flooding.
flooding does not occur. Our results are in good agreement with those Heavy rainfall usually occurs during the spring and fall, which are
of Khosravi et al. (2016a). more likely to produce high runoff, and it is at such times that hy­
Analyses of the flood susceptibility map (Fig. 6) indicate that 3% drologists and flood experts should implement plans to control future
and 12% of the entire country have very high and high susceptibility to floods, in particular for floods with high return periods. Flooding may
future flood occurrences, respectively, corresponding to total areas of occur in any region; thus, hydrologists should begin designing and
44,943 and 191,107 km2. These areas are distributed in northern and building flood control structures and applying management measures
western Iran (Fig. 6). The Guilan province has the country’s highest according to higher flood return periods. As our results indicate, and in
percentage of high and very high flood-prone areas (19.8%), followed agreement with well-established theoretical principles on flood pro­
by Mazandaran (17.6%), Golestan (10.1%), Alborz (10.2%), and Kho­ duction, surface water flows from mountain areas tend to concentrate
zestan (9.3%), whereas the Yazd province has the lowest percentage and produce flooding in the flat areas (valleys) located downstream
(0.2%; Fig. 7). By contrast, the highest percentage of areas with very (with lower slope and altitude). Even simple, practical measures can
low susceptibility to flooding is found in the Semnan province (51.4%), substantially reduce flood hazards and risks. To ensure that these
followed by Qom (36.4%) and Yazd (33.8%). This can be explained by measures can be implemented, the first step is to identify these flood-
the low amount of rainfall in these areas (i.e., deserts with an annual prone areas. The second step is to predict flood depths for different
rainfall of less than 200 mm). return periods, which will allow the delineation of critical areas. This
The susceptibility to flooding of a total of 1243 cities was analyzed combined information will facilitate the implementation of flood
in this study, as shown in Fig. 8. Around 29% and 49% of these cities measure plans according to the predicted flood depths and topo­
are located in areas with a high and very high probability of future graphical characteristics of the relevant area. It is important to note that
flood occurrences, respectively, and only 2% and 6% are located in different countries opt for different flood risk management portfolios
areas with very low and low hazards, respectively (Fig. 8). As Fig. 9 (Bubeck et al., 2017).
illustrates, the Mazandaran province, which has the second-highest A good example of efficient flood control structures is found in one
percentage of areas susceptible to flooding, has the highest percentage of the villages in Golestan province (Haji Abad). The residents of this
of cities (60 out of 67 cities, or 89.6%) exposed to high flood hazards, village, in anticipation of possible flooding based on warnings from
followed by Golestan (26 out of 31 cities, 83.9%), Charmahal and weather organizations, built a flood wall around the village using sacks
Bakhtiari (24 out of 30 cities, 80%), Lorestan (11 out of 14 cities, full of local soil, which has prevented inundations (Fig. 10). This ex­
78.6%), Khozestan (47 out of 62 cities, 75.8%), West Azarbayjan (17 ample shows that flooding can be managed using appropriate measures,
out of 24 cities, 70.8%), Hormozgan (17 out of 26 cities, 65.4%), and but control plans are unfortunately not available in all areas prone to
Kermanshah (18 out of 28 cities, 64.3). The Kerman province has one of flooding. Another major cause of flooding in Iran is the failure to re­
the lowest percentages of cities exposed to high flood hazards (20.8%), move alluvial depositional sediments and materials that settle in the
but it is still not a safe province, as 10 out of 48 cities are in serious beds of rivers in flat areas and cause the accretion of riverbeds, inducing
danger of future flooding. Overall, out of the 1243 cities considered, flooding for lower discharges (Hekmatifar et al., 2010). A lack of riv­
661 have very high susceptibility to future flooding hazards, 331 have erbed cleanup and vegetation removal from channels can also increase
high susceptibility, 157 have moderate susceptibility, and only 68 and flooding. Vegetation cover in the main river channel will increase the
26 have low and very low susceptibility, respectively. It is startling that flow’s roughness and, therefore, flood depths.
most of these cities were initially built in inappropriate locations and Although the severe 2019 floods in Iran corresponded to rainfall
have since expanded. In the past, the most important criterion for set­ events with long return periods and probably could not have been
tlement location was the proximity to rivers, but nowadays, because of prevented by riverbed soil removal, flood control structures and mea­
rapid changes in climate and anthropogenic disturbance as well as poor sures could have mitigated the extent of the flooding. Many cities in
catchment management, proximity to rivers has become one of the Iran, such as Ahvaz and Agh-ghala (Fig. 11), are located in close
greatest threats to human life. Our findings indicate that Iran faces proximity to large rivers, and therefore even rainfall events with lower
serious future flood hazards and that, consequently, flood mitigation return periods produce flooding, and entire cities can be quickly

9
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

inundated as a result of poor watershed management strategies. The 5. Of the 1243 cities considered in this study, 661 have very high
construction of flood wall control structures and/or detention dams and susceptibility to future flooding hazards, 331 have high suscept­
upland catchment management strategies could effectively prevent in­ ibility, 157 have moderate susceptibility, and only 68 and 26 have
undation in those cities. low and very low susceptibility, respectively.
A key weakness of flood susceptibility maps is that they lack in­ 6. The Mazandaran, Golestan, Charmahal and Bakhtiari, Lorestan,
formation regarding flood velocity, depth, and discharge. To implement Khozestan, West Azarbayjan, Hormozgan, and Kermanshah pro­
a comprehensive flood mitigation plan, it is important to focus on re­ vinces have the highest percentages of cities exposed to very high
gions with high and very high susceptibility to future flood occurrences future flooding risk.
and to identify critical regions from flood inundation maps, such as that 7. Although high rainfall is a key factor in flooding, poor watershed
produced as part of this study. Decision-makers can then follow up with management (as reflected by the effects of land use as the main
additional studies and field surveys to implement appropriate planning factor responsible for severe flooding) is the leading cause of
strategies to design effective prevention measures. Some of the best flooding events in Iran.
flood control measures include controlling runoff from upland areas by
planting vegetation, building check-dams and detention dams for return Findings from the present study will prove useful as a starting point
periods of more than 50 years, and constructing storage reservoirs in for further research on flood mitigation and can be used by the Iran
lowland areas and floodwalls in the periphery of urban communities as Water Resources Management Organization and decision-makers to
well as smaller dams and levees as appropriate. design appropriate watershed management strategies and plans. Flood
control measures in Iran are still nascent and should be prioritized to
5. Conclusion ensure that further flooding events are prevented.

Flood events are among the most destructive natural hazards and
Author contributions
cause irreparable damage worldwide each year. Because of its topo­
graphy and climate, Iran is prone to flooding and has experienced nu­
KK and MP came up with the idea, MP collected data and prepared
merous severe flooding events in recent years. The first essential step in
layers and performed CNN model. KK and AG did statistical analysis.
flood management and the generation of mitigation plans is to delineate
KK wrote the first drat. SDK, PMS, DTB, and SL provided critical in­
flood-prone areas clearly. In the present study, we prepared a flood
sights and contributed to the final version of the paper.
susceptibility map of Iran using a deep CNN algorithm. Ten flood
conditioning factors (slope, altitude, aspect, plan curvature, profile
curvature, rainfall, geology, land use, distance from roads, and distance Declaration of Competing Interest
from rivers) were identified and used in conjunction with data collected
from 2769 historical flood locations throughout the country to generate
The authors declare that they have no known competing financial
a flood susceptibility map. The resulting map was evaluated quantita­
interests or personal relationships that could have appeared to influ­
tively to identify areas with the highest and lowest flood risk and the
ence the work reported in this paper.
key factors contributing to this risk. Our main findings are as follows:

1. CNN has good and acceptable accuracy for identifying flood-prone Acknowledgments
areas.
2. Land use and geology are, respectively, the most and least effective This research was supported by the Basic Research Project of the
conditioning factors for flood occurrences in Iran. Korea Institute of Geoscience and Mineral Resources (KIGAM) and
3. Around 12% of Iran’s land is highly susceptible to flooding, and 3% Project of Environmental Business Big Data Platform and Center
is very highly susceptible. Construction funded by the Ministry of Science and ICT. Furthermore,
4. A total of 29% and 49% of Iran’s cities are at high and very high risk the authors would like to thank Iran water resources management for
for future flood hazards, respectively. providing the datasets.

10
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

Appendix

Fig. A1. Flood conditioning factors: (a) altitude, (b) slope degree, (c) aspect, (d) plan curvature, (e) profile curvature, (f) distance to river, (g) distance to road, (h)
rainfall, (i) geology, (j) land use.

11
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

Fig. A1. (continued)

12
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

Fig. B1. Non-flood locations in the country.

References 1019–1033.
Dehnavi, A., Aghdam, I.N., Pradhan, B., Varzandeh, M.H.M., 2015. A new hybrid model
using step-wise weight assessment ratio analysis (SWARA) technique and adaptive
Azareh, A., et al., 2019. Incorporating multi-criteria decision-making and fuzzy-value neuro-fuzzy inference system (ANFIS) for regional landslide hazard assessment in
functions for flood susceptibility assessment. Geocarto Int. 1–21. Iran. Catena 135, 122–148.
Bubeck, P., et al., 2017. Explaining differences in flood management approaches in Dodangeh, E., et al., 2020. Integrated machine learning methods with resampling algo­
Europe and in the USA–a comparative analysis. J. Flood Risk Manage. 10 (4), rithms for flood susceptibility prediction. Sci. Total Environ. 705, 135983.
436–445. Falah, F., et al., 2019. Artificial neural networks for flood susceptibility mapping in data-
Bui, D.T., et al., 2018. Novel hybrid evolutionary algorithms for spatial prediction of scarce urban areas, Spatial modeling in GIS and R for Earth and Environmental
floods. Sci. Rep. 8 (1), 1–14. Sciences. Elsevier, pp. 323-336.
Bui, D.T., et al., 2016. Hybrid artificial intelligence approach based on neural fuzzy in­ Fang, W., Ding, Y., Zhang, F., Sheng, V.S., 2019. DOG: A new background removal for
ference model and metaheuristic optimization for flood susceptibilitgy modeling in a object recognition from images. Neurocomputing 361, 85–91.
high-frequency tropical cyclone area using GIS. J. Hydrol. 540, 317–330. Ghorbani, M., 2013. Economic Geology of Iran, 581. Springer.
Chang, H.-S., Chen, T.-L., 2016. Spatial heterogeneity of local flood vulnerability in­ Ghorbanzadeh, O., Meena, S.R., Blaschke, T., Aryal, J., 2019. UAV-based slope failure
dicators within flood-prone areas in Taiwan. Environ. Earth Sci. 75 (23), 1484. detection using deep-learning convolutional neural networks. Remote Sensing 11
Chapi, K., et al., 2017. A novel hybrid artificial intelligence approach for flood suscept­ (17), 2046.
ibility assessment. Environ. Modell. Software 95, 229–245. Haghizadeh, A., Siahkamari, S., Haghiabi, A.H., Rahmati, O., 2017. Forecasting flood-
Chen, W., et al., 2019. Spatial prediction of groundwater potentiality using ANFIS en­ prone areas using Shannon’s entropy model. J. Earth Syst. Sci. 126 (3), 39.
sembled with teaching-learning-based and biogeography-based optimization. J. Hekmatifar, H., Nazariha, M., Givechi, S., 2010. Agricultural damage assessment due to
Hydrol. 572, 435–448. floods using HEC-RAS modeling and Arc View. J. Environ. Sci. Technol. 11 (4),
Chen, W., Panahi, M., Pourghasemi, H.R., 2017. Performance evaluation of GIS-based 95–109.
new ensemble data mining techniques of adaptive neuro-fuzzy inference system Hens, L., et al., 2018. Sea-level rise and resilience in Vietnam and the Asia-Pacific: a
(ANFIS) with genetic algorithm (GA), differential evolution (DE), and particle swarm synthesis. Vietnam J. Earth Sci. 40 (2), 126–152.
optimization (PSO) for landslide spatial modelling. Catena 157, 310–324. Hirabayashi, Y., et al., 2013. Global flood risk under climate change. Nat. Clim. Change 3
Chen, Y.-R., Yeh, C.-H., Yu, B., 2011. Integrated application of the analytic hierarchy (9), 816–821.
process and the geographic information system for flood risk assessment and flood Hou, J., Lu, Y., Li, Z., Zhu, H., 2020. Effects of different functional structure parameters of
plain management in Taiwan. Nat. Hazards 59 (3), 1261–1276. plant communities on slope runoff in different periods of the year in semiarid
Choubin, B., et al., 2019. An ensemble prediction of flood susceptibility using multi­ grasslands. Sci. Total Environ. 713, 136705.
variate discriminant analysis, classification and regression trees, and support vector Kavzoglu, T., Sahin, E.K., Colkesen, I., 2014. Landslide susceptibility mapping using GIS-
machines. Sci. Total Environ. 651, 2087–2096. based multi-criteria decision analysis, support vector machines, and logistic regres­
Chung, C.-J.-F., Fabbri, A.G., 2003. Validation of spatial prediction models for landslide sion. Landslides 11 (3), 425–439.
hazard mapping. Nat. Hazards 30 (3), 451–472. Khosravi, K., Nohani, E., Maroufinia, E., Pourghasemi, H.R., 2016a. A GIS-based flood
Dano, U.L., et al., 2019. Flood susceptibility mapping using GIS-based analytic network susceptibility assessment and its mapping in Iran: a comparison between frequency
process: a case study of Perlis, Malaysia. Water 11 (3), 615. ratio and weights-of-evidence bivariate statistical models with multi-criteria deci­
Das, S., 2019. Geospatial mapping of flood susceptibility and hydro-geomorphic response sion-making technique. Nat. Hazards 83 (2), 947–987.
to the floods in Ulhas basin, India. Remote Sens. Appl.: Soc. Environ. 14, 60–74. Khosravi, K., Panahi, M., Tien Bui, D., 2018a. Spatial prediction of groundwater spring
de Brito, M.M., Evers, M., 2016. Multi-criteria decision-making for flood risk manage­ potential mapping based on an adaptive neuro-fuzzy inference system and meta­
ment: a survey of the current state of the art. Nat. Hazards Earth Syst. Sci 16 (4), heuristic optimization. Hydrol. Earth Syst. Sci. 22 (9).

13
K. Khosravi, et al. Journal of Hydrology 591 (2020) 125552

Khosravi, K., et al., 2018b. A comparative assessment of decision trees algorithms for Smith, K., 2003. Environmental Hazards: Assessing Risk and Reducing Disaster.
flash flood susceptibility modeling at Haraz watershed, northern Iran. Sci. Total Routledge.
Environ. 627, 744–755. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R., 2014. Dropout:
Khosravi, K., Pourghasemi, H.R., Chapi, K., Bahri, M., 2016b. Flash flood susceptibility a simple way to prevent neural networks from overfitting. J. Machine Learning Res.
analysis and its mapping using different bivariate models in Iran: a comparison be­ 15 (1), 1929–1958.
tween Shannon’s entropy, statistical index, and weighting factor models. Environ. Tehrany, M.S., Pradhan, B., Jebur, M.N., 2013. Spatial prediction of flood susceptible
Monit. Assess. 188 (12), 656. areas using rule based decision tree (DT) and a novel ensemble bivariate and mul­
Khosravi, K., et al., 2019. A comparative assessment of flood susceptibility modeling tivariate statistical models in GIS. J. Hydrol. 504, 69–79.
using Multi-Criteria Decision-Making Analysis and Machine Learning Methods. J. Tehrany, M.S., Pradhan, B., Jebur, M.N., 2014. Flood susceptibility mapping using a novel
Hydrol. 573, 311–323. ensemble weights-of-evidence and support vector machine models in GIS. J. Hydrol.
Kia, M.B., et al., 2012. An artificial neural network model for flood simulation using GIS: 512, 332–343.
Johor River Basin, Malaysia. Environ. Earth Sci. 67 (1), 251–264. Tehrany, M.S., Pradhan, B., Jebur, M.N., 2015. Flood susceptibility analysis and its ver­
Kourgialas, N.N., Karatzas, G.P., 2011. Flood management and a GIS modelling method to ification using a novel ensemble support vector machine and frequency ratio method.
assess flood-hazard areas—a case study. Hydrol. Sci. J.-Journal des Sciences Stoch. Env. Res. Risk Assess. 29 (4), 1149–1165.
Hydrologiques 56 (2), 212–225. Teng, J., et al., 2017. Flood inundation modelling: a review of methods, recent advances
LeCun, Y., Bottou, L., Bengio, Y., Haffner, P., 1998. Gradient-based learning applied to and uncertainty analysis. Environ. Modell. Software 90, 201–216.
document recognition. Proc. IEEE 86 (11), 2278–2324. Termeh, S.V.R., Kornejady, A., Pourghasemi, H.R., Keesstra, S., 2018. Flood susceptibility
Lee, M.-J., Kang, J.-e., Jeon, S., 2012. Application of frequency ratio model and validation mapping using novel ensembles of adaptive neuro fuzzy inference system and me­
for predictive flooded area susceptibility mapping using GIS, 2012 IEEE international taheuristic algorithms. Sci. Total Environ. 615, 438–451.
geoscience and remote sensing symposium. IEEE, pp. 895-898. Tien Bui, D., et al., 2018. New hybrids of anfis with several optimization algorithms for
Lu, H., et al., 2017. Cultivated land information extraction in UAV imagery based on deep flood susceptibility modeling. Water 10 (9), 1210.
convolutional neural network and transfer learning. J. Mountain Sci. 14 (4), Tien Bui, D., et al., 2019. Flood spatial modeling in northern Iran using remote sensing
731–741. and gis: a comparison between evidential belief functions and its ensemble with a
Panahi, M., Sadhasivam, N., Pourghasemi, H.R., Rezaie, F., Lee, S., 2020. Spatial pre­ multivariate logistic regression model. Remote Sensing 11 (13), 1589.
diction of groundwater potential mapping based on convolutional neural network Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., 2012. Landslide susceptibility as­
(CNN) and support vector regression (SVR). J. Hydrol. 125033. sessment in vietnam using support vector machines, decision tree, and Naive Bayes
Pappenberger, F., Matgen, P., Beven, K.J., Henry, J.-B., Pfister, L., 2006. Influence of Models. Mathematical problems in Engineering, 2012.
uncertain boundary conditions and model structure on flood inundation predictions. Tiryaki, M., Karaca, O., 2018. Flood susceptibility mapping using GIS and multicriteria
Adv. Water Resour. 29 (10), 1430–1449. decision analysis: Saricay-Çanakkale (Turkey). Arabian J. Geosci. 11 (14), 364.
Pham, T.D., Bui, D.T., Yoshino, K., Le, N.N., 2018. Optimized rule-based logistic model Van Thanh, N., Le, D.T., Thinh, N.A., Lan, T.D., Hens, L., 2017. Shifting challenges for
tree algorithm for mapping mangrove species using ALOS PALSAR imagery and GIS coastal green cities. Vietnam J. Earth Sci. 39 (2), 109–129.
in the tropical region. Environ. Earth Sci. 77 (5), 159. Wang, Y., Fang, Z., Hong, H., 2019. Comparison of convolutional neural networks for
Rahmati, O., Pourghasemi, H.R., Zeinivand, H., 2016. Flood susceptibility mapping using landslide susceptibility mapping in Yanshan County, China. Sci. Total Environ. 666,
frequency ratio and weights-of-evidence models in the Golastan Province, Iran. 975–993.
Geocarto Int. 31 (1), 42–70. Yadollahie, M., 2019. The flood in Iran: a consequence of the global warming? Int. J.
Sahoo, G., Schladow, S., Reuter, J., 2009. Forecasting stream water temperature using Occupational Environ. Med. 10 (2), 54.
regression analysis, artificial neural network, and chaotic non-linear dynamic models. Yesilnacar, E.K., 2005. The application of computational intelligence to landslide sus­
J. Hydrol. 378 (3–4), 325–342. ceptibility mapping in Turkey. University of Melbourne, Department, 200.
Sarhadi, A., Soltani, S., Modarres, R., 2012. Probabilistic flood inundation mapping of Youssef, A.M., Pradhan, B., Pourghasemi, H.R., Abdullahi, S., 2015. Landslide suscept­
ungauged rivers: Linking GIS techniques and frequency analysis. J. Hydrol. 458, ibility assessment at Wadi Jawrah Basin, Jizan region, Saudi Arabia using two bi­
68–86. variate models in GIS. Geosci. J. 19 (3), 449–469.

14

You might also like