Deep Learning Model and Ensemble Multi-Criteria

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Journal of Environmental Management 345 (2023) 118838

Contents lists available at ScienceDirect

Journal of Environmental Management


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

Research article

Novel integrated modelling based on multiplicative long short-term


memory (mLSTM) deep learning model and ensemble multi-criteria
decision making (MCDM) models for mapping flood risk
Aliakbar Mohammadifar a, Hamid Gholami a, *, Shahram Golzari b, c
a
Department of Natural Resources Engineering, University of Hormozgan, Bandar-Abbas, Hormozgan, Iran
b
Department of Electrical and Computer Engineering, University of Hormozgan, Bandar-Abbas, Hormozgan, Iran
c
Deep Learning Research Group, University of Hormozgan, Bandar-Abbas, Hormozgan, Iran

A R T I C L E I N F O A B S T R A C T

Handling Editor: Jason Michael Evans Flood risk assessment is a key step in flood management and mitigation, and flood risk maps provide a quan­
titative measure of flood risk. Therefore, integration of deep learning – an updated version of machine learning
Keywords: techniques – and multi-criteria decision making (MCDM) models can generate high-resolution flood risk maps. In
Deep learning this study, a novel integrated approach has been developed based on multiplicative long short-term memory
Feature selection
(mLSTM) deep learning models and an MCDM ensemble model to map flood risk in the Minab-Shamil plain,
Flood risk map
southern Iran. A flood hazard map generated by the mLSTM model is based on nine critical features selected by
CODAS-EDAS-MOOSRA ensemble
Iran GrootCV (distance to the river, vegetation cover, variables extracted from DEM (digital elevation model) and
river density) and a flood inventory map (70% and 30% data were randomly selected as training and test
datasets, respectively). The values of all criteria used to assess model accuracy performance (except Cohens
kappa for train dataset = 86, and for test dataset = 84) achieved values greater than 90, which indicates that the
mLSTM model performed very well for the generation of a spatial flood hazard map. According to the spatial
flood hazard map produced by mLSTM, the very low, low, moderate, high and very high classes cover 26%,
35.3%, 20.5%, 11.2% and 7% of the total area, respectively. Flood vulnerability maps were produced by the
combinative distance-based assessment (CODAS), the evaluation based on distance from average solution
(EDAS), and the multi-objective optimization on the basis of simple ratio analysis (MOOSRA), and then validated
by Spearman’s rank correlation coefficients (SRC). Based on the SRC, the three models CODAS, EDAS, and
MOOSRA showed high-ranking correlations with each other, and all three models were then used in the
ensemble process. According to the CODAS-EDAS-MOOSRA ensemble model, 21.5%, 34.2%, 23.7%, 13%, and
7.6% of the total area were classified as having a very low to very high flood vulnerability, respectively. Finally, a
flood risk map was generated by the combination of flood hazard and vulnerability maps produced by the
mLSTM and MCDM ensemble model. According to the flood risk map, 27.4%, 34.3%, 14.8%, 15.7%, and 7.8% of
the total area were classified as having a very low, low, moderate, high, and very high flood risk, respectively.
Overall, the integration of mLSTM and the MCDM ensemble is a promising tool for generating precise flood risk
maps and provides a useful reference for flood risk management.

1. Introduction 2007; Ekwueme and Agunwamba, 2021). Annually, 21 million people


are impacted by floods worldwide, this being expected to reach fiftyfour
Floods – as natural multidimensional hydrological hazard phenom­ million by 2030 (Kundzewicz, 2002). Although it is impossible to
ena (Tsakiris et al., 2015) – result from heavy rainfall in combination minimize flood risk in a river catchment to zero, it is of utmost impor­
with complex urban infrastructure, agricultural practices, roads, reser­ tance to decrease the consequences of flood hazards to a minimum.
voirs, human dwelling habits, and increasing sea-level rise, thereby Flood assessment is necessary for a successful management of such
causing land instability such as subsidence and soil erosion (Parry et al., events. It aims at identifying the areas of highest risk and at mitigating

* Corresponding author.
E-mail address: hgholami@hormozgan.ac.ir (H. Gholami).

https://doi.org/10.1016/j.jenvman.2023.118838
Received 23 June 2023; Received in revised form 30 July 2023; Accepted 14 August 2023
Available online 16 August 2023
0301-4797/© 2023 Elsevier Ltd. All rights reserved.
A. Mohammadifar et al. Journal of Environmental Management 345 (2023) 118838

its detrimental economic, social, and environmental consequences. Due Iran experienced more than 3700 flood events over the past 60 years
to recent advancements in computer science, many ML algorithms have (Watershed Department of Iran, 2002). Due to the increasing urbani­
been employed for flood assessment in recent years (Chen et al., 2021; zation rate (Yari et al., 2019), land use changes, climate change, and
Ekwueme, 2022). The application of computer-based learning algo­ reservoir mismanagement, such flood events have increasingly affected
rithms, especially deep learning (DL) models (e.g., graph convolutional local populations, causing economic losses of >1.7 billion $ US (Khos­
networks (GCN, Gholami et al., 2023), bidirectional gated recurrent unit ravi et al., 2016; Norouzi and Taslimi, 2012).
(BiGRU; Rezaei et al., 2023), stacking- and voting-based ensemble DL In January 2022, the Hormozgan province of southern Iran experi­
(Mohammadifar et al., 2023), hybrid DL (Gholami and Mohammadifar, enced heavy rainfalls, in the course of which Minab city located east of
2022), gcForest and bidirectional long short-term memory (BiLSTM; the capital of Hormozgan province – Bandar Abbas – experienced two
Gholami et al., 2021a), restricted Boltzmann machine (Gholami et al., destructive floods, one on 4 January and another on January 18, 2022.
2021b), and etc.), are being increasingly applied to the spatial modelling Based on the records of the Minab meteorological station (with a mean
of different environmental hazards (e.g., soil erosion by water and wind, annual rainfall of 162 mm), the Minab River catchment, in which the
dust emissions, land subsidence, landslide, soil salinization), especially Esteghlal Dam is located, received more than 400 mm of rainfall in
flood hazard, vulnerability, and risk assessment (Costache et al., 2022b). January 2022 alone. The main consequences of the two flood events on
For example, some studies (e.g., Wu et al., 2020; Panahi et al., 2021; the Minab-Shamil plain were severe economic losses resulting from the
Costache et al., 2022b) applied DL models (e.g., RNN, CNN, LSTM, destruction of roads and bridges. Furthermore, intensified soil erosion in
GBDT, and DBN) for the spatial and temporal prediction of a flood, its the upstream watershed of the Minab River led to an increased sus­
depth and susceptibility. pended sediment load in the river networks, urban flooding, and
MCDM is a helpful technique for solving sophisticated decision impaired city services. In view of this situation, the prediction and
challenges required for non-comparable criteria and data (Malczewski, management of catastrophic flood events has become a high priority in
2006). Individual MCDM models (e.g., technique for order preference by Iran.
similarity to ideal solution (TOPSIS), fuzzy analytic hierarchy process To mitigate detrimental flood effects in the future, the main objec­
(FAHP), decision-making trial and evaluation laboratory (DEMATEL), tives of this study are: (i) to map flood hazard spatially by means of the
elimination and choice expressing reality (ELECTRE), etc.) or their mLSTM model, (ii) to map flood vulnerability by the CODAS-EDAS-
combination with GIS, statistical methods and data mining models have MOOSRA ensemble model, and (iii) to map flood risk by the mLSTM
been used to assess different aspects of flood management (e.g., Arora and MCDM ensemble.
et al., 2021; Costache et al., 2020; Costache et al., 2022a,b; Arabameri
et al., 2019, 2020). However, so far – and based on our knowledge – 2. Materials and methods
application of EDAS, MOOSRA, and CODAS have not previously been
reported. Therefore, in the present study, three MCDM models (e.g., 2.1. Study area
EDAS, MOOSRA, and CODAS) for assessing flood vulnerability were
applied for the first time. The Minab-Shamil plain (27◦ to 28◦ N and 56◦ 20‘ to 57◦ 20‘ E) with

Fig. 1. Location of the study area with the flood inventory map indicating the spatial distribution of flood points in the region.

2
A. Mohammadifar et al. Journal of Environmental Management 345 (2023) 118838

an area of 4762 km2 is located on the northern coast of the Persian Gulf, adding connections from the mRNN’s intermediate state mt to each
and in the eastern part of the Hormozgan province, southern Iran gating unit in the LSTM, resulting in the following system:
(Fig. 1). Based on long-term data from the Minab meteorological station,
mt = (Wmxxt) ʘ (Wmhht-1) (2)
the mean annual precipitation is 162 mm, and the mean annual tem­
perature 26.9 ◦ C. The study area is the most fertile plain in the Hor­ ̂ t = Whx xt + Whm mt
h (3)
mozgan province, and the dominant use of its land is for agricultural
production. Based on a digital elevation model (DEM), elevations range it = σ (Wixxt + Wimmt) (4)
from − 20 m (regions located along the northern coast of the Persian
Gulf) to 3120 m. The main river in the study area is the Minab with ot = σ (Woxxt + Wommt) (5)
perennial flow. The Minab Dam established on the Minab River supplies
ft = σ (Wfxxt + Wfmmt) (6)
domestic water to the Minab and Bandar Abbas cities, irrigation water,
artificial recharge of groundwater and flood mitigation. where xt and ht-1 are the input layer and the previous hidden state,
respectively; σ is the logistic sigmoid function; h
̂ t indicates the input to
2.2. Methods the next hidden state before any non-linear operation; it denotes the
input gate; ot and ft are the output gate and forget gate, respectively.
2.2.1. Mapping flood hazard by mLSTM mLSTM is available in the github package Python. The structure of
our model has two layers consisting of mLSTM (with unit number = 128,
2.2.1.1. Flood hazard conditioning variables and its inventory map. The dropout = 0.2, and recurrent dropout = 0.2), and dense (with unit
variables used to assess flood hazard include a digital elevation model number = 1, and the sigmoid activation function). At the model
(DEM), slope, topographic curvature, plan, profile, topographic wetness compilation stage, the binary cross-entropy and accuracy were selected
index (TWI), distance to river, vertical distance to channel, river density, as loss function and metrics, respectively. In comparison with the Ada­
topographic position index (TPI), stream power index (SPI), normalized max optimizer, the model with the Adam optimizer had higher accuracy.
difference vegetation index (NDVI), soil adjusted vegetation index At the fit stage, the model was tested with different epoch numbers (500,
(SAVI), land use, lithology, soil, and a flood inventory map. Shuttle 700, and 1000) and batch sizes (10, 15, and 20), of which the model
radar topography mission (SRTM) images with a resolution of 90 × 90 m with the batch size = 10 and epoch = 500 were the most accurate. The
were downloaded from https://earthexplorer.usgs.gov, and then used to structure of mLSTM is presented in Fig. 2.
generate the DEM. Based on the DEM, the slope generated by the spatial Six criteria consisting of accuracy, precision, recall, F1 score,
analysis tool in the ArcGIS10.3 software package and six variables Cohen’s kappa, and the receiver operating characteristic - area under
consisting of topographic curvature, plan, profile, TWI, TPI, and SPI curve (ROC-AUC) were applied to assess the flood hazard map generated
were generated in the SAGA-GIS. by the mLSTM model.

TWI = Ln (As/tanθ) (1) 2.2.2. Mapping flood vulnerability by MCDM


2
where As and θ indicate the specific area (m ) and slope (degree),
2.2.2.1. Flood vulnerability conditioning variables. The variables used to
respectively.
map the flood vulnerability are: (1) The population density, villages, and
The variables related to the river (e.g., distance to the river, vertical
residential regions density (related data received from the Statistical
distance to channel, and river density) were generated by different
center of Iran), and then three layers consisting of the population den­
functions in Arc-hydro. For producing NDVI and SAVI, the Landsat im­
sity, villages, and residential area density were mapped in ArcGIS. (2)
ages recorded on October 25, 2020 were downloaded, and then trans­
Distance to roads (the newest data were downloaded from www.ope
formed into a mosaic with the ENVI software. Three maps consisting of
nstreetmap.org), and then the “distance” function was applied in Arc­
land use, lithology, and soil were generated based on maps produced by
GIS to calculate the distance to roads. (3) Distance to medical centers
the Iran Forests, Rangelands, and Watershed Management Organization
(location of the medical centers was extracted from the OSM maps), and
(IFRWMO). All nine important features are presented in suppl. 1 and 2.
then the “distance” function was applied in the ArcGIS to calculate the
A flood inventory map (Fig. 1) – a map showing the spatial distri­
distance to medical centers. (4) Distance to the livestock, tourism,
bution of floods in the study area – is essential for constructing predic­
mineral, and industrial centers (these layers were generated based on
tive models, and this map has a very strong influence on the accuracy of
the Land Management studies for Hormozgan province). (5) Distance to
the flood hazard map. It is a key step to explore the spatial relation
the power and energy transmission lines (this variable is extracted from
between flood occurrences and control variables. Here, a flood in­
the OSM maps). The spatial maps for factors used to map flood vulner­
ventory map was prepared based on a field survey, historical flood signs,
ability are presented in suppl. 3 and 4.
and a map produced by IFRWMO. For constructing the flood hazard
predictive model, we randomly selected 70% (92 flood points) and 30%
(39 flood points) of the data as training and test datasets, respectively. 2.2.2.2. Calculating weights of flood vulnerability conditioning factors. To
calculate the weight and importance of the flood vulnerability condi­
tioning variables, we applied the analytic hierarchy process (AHP).
2.2.1.2. Determining importance variables related to the flood hazard by
GrootCV. Feature selection is the process of reducing input data (Nilsson
et al., 2007) and represents an important step of building a predictive 2.2.2.3. MCDM models. Three MCDM models consisting of CODAS,
model based on the machine learning algorithms (Tan and He, 2021). EDAS and MOOSRA were applied to map flood vulnerability.
Here, we apply GrootCV to select the most important variables con­ CODAS – suggested by Ghorabaee et al. (2017) – is a new MCDM
trolling the flood. model that has some features not considered in the other MCDM models.
It uses the Euclidean and Texicab distances from the negative-ideal
point. EDAS – introduced initially by Ghorabaee et al. (2015b) – is a
2.2.1.3. The mLSTM deep learning model for spatial flood assessment. The
helpful technique to handle MCDM problems (Darko and Liang, 2020).
LSTM network as a type of RNN has three gates consisting of input (i),
Unlike other MCDM models (e.g., TODIM, VIKOR, and TOPSIS), EDAS
forget (f), and output (o) (Hochreiter and Schmidhuber, 1997; Krause
uses the average solutions for evaluating the alternatives.
et al., 2017). mLSTM – a hybrid architecture – combines LSTM and
MOOSRA as a multi-objective optimization technique (Aytaç Adalı
multiplicative RNN (mRNN) for sequence modeling (Krause et al., 2017;
and Tuş Işık, 2017) uses operator ‘division’ instead of ‘subtraction’.
Ma et al., 2021). The mRNN and LSTM architectures can be combined by

3
A. Mohammadifar et al. Journal of Environmental Management 345 (2023) 118838

Fig. 2. Conceptual diagram of mLSTM for predicting flood hazard.

Unlike other MCDM models (e.g., VIKOR, MABAC, and GRA), MOOSRA vulnerability predicted by mLSTM and the CODAS-EDAS-MOOSRA
chooses the best alternative with high simplification, low computational ensemble, respectively. All key steps used to map flood risk are pre­
time, ignores the extra variables used in other models mentioned above, sented in Fig. 3.
and carries out basic mathematical calculations (Narayanamoorthy
et al., 2020). 3. Results and discissuion

2.2.2.4. Ensemble MCDM models. The Spearman’s rank correlation co­ 3.1. Discriminating important features controlling flood hazard from non-
efficient (R) is applied to measure the degree of similarity between important ones
ranking pairs of different MCDM models for determining the best MCDM
model (Rao, 2007; Athawale and Chakraborty, 2011). R can be Fig. 4 shows the important (confirmed) and non-important (rejected)
described as: features controlling flood hazard identified by GrootCV. Based on the
∑ results, among 16 features (variables) explored as flood hazard con­
6 mi=1 di2
R=1 − (8) trollers, nine features comprising distance to the river, NDVI, elevation
m (m2 − 1)
(DEM), SAVI, slope, TPI, SPI, river density, and TWI were selected as
where di and m indicate the difference between ranks in two different important variables, and seven variables consisting of land use, VDC,
MCDM models for the same alternative and the total number of alter­ soil order, profile, Lithology, curvature and plan were selected as non-
natives, respectively. important variables and were excluded from further analysis. Eleva­
The MCDM models with a high correlation coefficient and signifi­ tion derived from the DEM, slope, NDVI, and distance to river networks
cance level (p ≤ 0.01) were then used for the ensemble modeling have previously been identified as important features controlling flood
process. events (Pradhan, 2010). Flat regions of low elevation and steeper slopes
are particularly susceptible areas to flood occurrence. Slope as a key
2.2.3. Flood risk map flood variable decreases the infiltration time and, as a consequence,
Some studies (e.g., Winsemius et al., 2013; Jongman et al., 2012; causes runoff to increase enormously (Tehrany et al., 2015). Distance to
Pham et al., 2021) have used flood hazard, vulnerability, and exposure the river is the most important feature controlling flood hazard in the
for assessing flood risk, whereas others (e.g., Masuya, 2014; Scheuer study area because it represents the area where water is most concen­
et al., 2011) have combined two flood components, hazard and trated (Fernández and Lutz, 2010). Coupled with river density it rep­
vulnerability, for flood risk assessment. Here, we have combined flood resents the area adjacent to the river that is most susceptible to flooding
hazard with flood vulnerability for assessing flood risk. Flood risk (FR) is (Glenn et al., 2012). According to Ngo et al. (2021) and Pham et al.
a 2-dimensional index consisting of flood hazard (H) and vulnerability (2021), elevation, slope, and land cover are the most important in­
(V). FR is mapped by an integrated modeling approach based on the dicators of flood susceptibility. SPI and TWI are important indicators for
mLSTM and MCDM ensemble models: the erosive power in overland flow. As topographic variables they
control a variety of different hydrological phenomena associated with
FR = HmLSTM * VCODAS-EDAS-MOOSRA ensemble (9) flood occurrence (Moore and Grayson, 1991). Indeed, slope, distance to
the river, and land use/land cover have previously been identified as
where HmLSTM and VCODAS-EDAS-MOOSRA ensemble are the flood hazard and
critical variables controlling flood events in Northern Iran by Arabameri

4
A. Mohammadifar et al. Journal of Environmental Management 345 (2023) 118838

Fig. 3. Key steps used for the spatiotemporal assessment of the different aspects of a flood. DR and RD indicate distance from river and river density, respectively.

et al. (2019). susceptible regions to flood hazard.


The results of six statistical criteria used to assess the model’s per­
3.2. The flood hazard map and its performance formance for the train and test datasets are presented in Table 1. The
values of all criteria (except Cohens kappa for train dataset = 86, and for
The spatial map of flood hazard generated by mLSTM is presented in test dataset = 84) for the two datasets were calculated at > 90, which
Fig. 5. Flood hazard is divided into five classes comprising very low indicates that the mLSTM model is highly accurate in mapping flood
(0–0.2), low (0.2–0.4), moderate (0.4–0.6), high (0.6–0.8), and very hazard. Based on Yesilnacar (2005), mLSTM (with ROC-AUC = 93% for
high (0.8–1) categories. According to the spatial flood hazard map the training dataset, and ROC-AUC = 92% for the test dataset) was
produced by mLSTM, the very low, low, and moderate classes cover judged to be of excellent performance. Despite the high performance of
1232 km2 (26%), 1682 km2 (35.3%), and 973 km2 (20.5%), respec­ mLSTM (with AUC >90% indicating an excellent performance) in pre­
tively, whereas 539 km2 (11.2%) and 335 km2 (7%) of the total area dicting flood hazard, we are not aware of any report indicating the
belong to the high and very high hazard class, respectively. Overall, the application of this model for flood hazard assessment, but there are a few
areas located in the vicinity of major rivers and streams are the most studies on the application of LSTM for flood modeling. For example,

5
A. Mohammadifar et al. Journal of Environmental Management 345 (2023) 118838

Fig. 4. Important (confirmed) and non-important (rejected) features identified by the GrootCV feature selection algorithm.

Fig. 5. Flood hazard map produced by mLSTM.

some researchers (e.g., Luppichini et al., 2022) reported the successful


Table 1 application of LSTM for flood prediction and its susceptibility. In com­
The results of the statistical criteria for assessment of mLSTM model
parison with conventional RNN, LSTM has a better performance in
performance.
solving sequential tasks (Hochreiter and Schmidhuber, 1997). Conven­
Criteria Train dataset Test dataset tional RNN has the problems of gradient vanishing and exploding, and is
Accuracy (%) 93 92 also challenging in solving the long-term sequential input problem
Precision (%) 96 95 (Bengio et al., 1994). In another study by Ramayanti et al. (2022), an
Recall (%) 93 92
integrated model based on synthetic aperture radar images, and two DL
F1 score (%) 95 94
Cohens kappa (%) 86 84 models (CNN and the group method of data handling (GMDH)) were
ROC AUC (%) 93 92 introduced to map flood susceptibility. Their findings showed that the

6
A. Mohammadifar et al. Journal of Environmental Management 345 (2023) 118838

CNN model performed better than the GMDH model. In comparison with
integrated models (e.g., ICO-ADT-FR and ICO-MLP-FR), predictive
models based on integration of the iterative classifier optimizer, deep
learning neural network and frequency ratio exhibit higher performance
in mapping flood susceptibility (Costache et al., 2022b). Flash-flood
susceptibility generated by a deep learning neural network – an
analytical hierarchy process (DLNN-AHP) – is more accurate than those
produced by MLP-AHP, fuzzy AHP and naïve bayes-AHP models (Cos­
tache et al., 2022a). Overall, our findings about the flood hazard can be
useful for the mitigation of soil erosion by water and its negative con­
sequences such as riverine sediment load threating key ecosystem
services.

3.3. The flood vulnerability map

3.3.1. Weights of flood vulnerability conditioning variables


The weights of variables control flood vulnerability calculated by
AHP (Fig. 6). Based on the results, the weights of flood vulnerability
conditioning variables assessed by AHP were as follows: population
density > village density > distance to livestock center > distance to
mining and industrial centers > distance to medicine center > distance
to road > distance to tourism center > distance to power and energy
transmission lines.

3.3.2. Flood vulnerability maps generated by three individual MCDM


models
The results of the three MCDM models used to map the flood
vulnerability in the study area are presented in Fig. 7 and Table 2. Based
on the CODAS model, 23%, 52.5%, 9.5%, 7.5%, and 7.5% of the total
area were classified from very low to very high in flood vulnerability,
respectively. The map generated by the EDAS model suggested that
5.5%, 17.5%, and 47.5% of the total area were classified as very low to
moderate in flood vulnerability, whereas 23.5% and 6% of the study
area were classified as high and very high in vulnerability, respectively.
The results of the MOOSRA model predicted that 39.2% and 40.5% of
the area belonged to the very low and low classes in flood vulnerability,
whereas 15.6%, 4.2%, and 0.5% of the total area were classified as
moderate, high, and very high in flood susceptibility, respectively. Due
to the high density of roads, power transmission lines, proximity of the
population, tourism, services, and economic centers to river networks,
the lands located downstream of the Esteghlal Dam were classified as
high and very high in flood vulnerability, whereas the northern parts of
the study area were classified as being of low and very low vulnerability.
Therefore, the results of MCDM models for assessing flood vulnerability
can be helpful in mitigating economic losses due to damage to basic and

Fig. 7. Flood vulnerability maps generated by (a) CODAS, (b) EDAS, and
(c) MOOSRA.

key infrastructures such as bridges, roads and energy lines.

3.3.3. Flood vulnerability map generated by the CODAS-EDAS-MOOSRA


ensemble
Fig. 6. Weights of variables that control flood vulnerability as calculated by Based on the Spearman’s rank correlation coefficient (R), all three
AHP. PD, VD, DL, DMI, DM, DR, DT, and DP indicate population density, village MCDM models (CODAS, EDAS, and MOOSRA) were correlated with a
density, distance to livestock centers, distance to mining and industrial centers, significance level of 99% (p ≤ 0.01) (Table 3). It is also noted that the
distance to medical centers, distance to roads, distance to tourism centers, and rankings obtained from the three MCDM models are identical. Based on
distance to power and energy transmission lines, respectively.

7
A. Mohammadifar et al. Journal of Environmental Management 345 (2023) 118838

Table 2 are suitable for ensemble modeling. According to the CODAS-EDAS-


Areas of flood vulnerability classes calculated by three MCDM models. MOOSRA ensemble (Fig. 8), 1027 km2 (21.5%), 1623 km2 (34.2%),
Vulnerability MCDM model 1127 km2 (23.7%), 620 km2 (13%), and 361 km2 (7.6%) of the total area
class were classified as very low to very high in flood vulnerability,
CODAS EDAS MOOSRA
respectively.
Area % Area % Area %
(km2) (km2) (km2)
3.4. Flood risk map generated by integrating mLSTM with the CODAS-
Very low 1088 23 264 5.5 1870 39.2
Low 2501.6 52.5 837 17.5 1923 40.5
EDAS-MOOSRA ensemble
Moderate 451 9.5 2266 47.5 743 15.6
High 356 7.5 1108 23.5 200 4.2 Iran, especially its southern parts, is affected by floods every year and
Very high 362 7.5 286 6 26 0.5 hence requires effective tools for risk assessment and management.
Therefore, an accurate map of flood risk is not only necessary for the
sustainable management of the watershed (Ngo et al., 2021), but also for
Table 3 the mitigation of negative flood consequences. To date, we are not aware
Cross-correlation coefficients obtained by three MCDM models. of any applications of the mLSTM and MCDM models to map flood risk.
Correlations In the present study, we have integrated the mLSTM and ensemble
MCDM models to assess the risk of floods in the study area. The resulting
CODAS EDAS MOOSRA
spatial map of flood risk generated in this way is presented in Fig. 9. The
a
CODAS Correlation coefficient 1.000 .828 .689a integrated model (mLSTM-CODAS-EDAS-MOOSRA) predicted that
Sig. (2-tailed) . .000 .000
N 9634 9634 9634
1301 km2 (27.4%), 1634 km2 (34.4%), 703 km2 (14.8%), 745 km2
EDAS Correlation coefficient .828a 1.000 .949a (15.7%) and 372 km2 (7.7%) of the total area can be classified as very
Sig. (2-tailed) .000 . .000 low to very high in flood risk, respectively. Based on the integrated
N 9634 9634 9634 model, the regions located downstream of the Esteghlal Dam (southern
MOOSRA Correlation coefficient .689a .949a 1.000
and southeastern parts of the study area having a flat topography, low
Sig. (2-tailed) .000 .000 .
N 9634 9634 9634 elevation, the highest density of population, residential and industrial
a
centers, and with intense agricultural and residential uses) were classi­
Correlation is significant at the 0.01 level (2-tailed).
fied as high and very high in flood risk. By contrast, the northern parts of
the study area characterized by mountainous landscapes, high elevation,
Table 3, all three models are well correlated with each other. Thus, complicated topography, and low density of villages and population
CODAS is highly positively correlated with EDAS (R = 0.828) and were classified as low and very low in flood risk. Integration of deep
MOOSRA (R = 0.689). It is also noted that EDAS has the highest (pos­ neural networks (DNNs) – a DL model – with MCDM models is thus a
itive) correlation with MOOSRA (R = 0.949). Therefore, all three models promising tool for generating accurate maps of flood risk for better flood

Fig. 8. Flood vulnerability map generated by the CODAS-EDAS-MOOSRA ensemble.

8
A. Mohammadifar et al. Journal of Environmental Management 345 (2023) 118838

Fig. 9. Flood risk map generated by integrating mLSTM with the CODAS-EDAS-MOOSRA ensemble.

management (Pham et al., 2021). Integration of machine learning sedimentation. Therefore, the results of this study can be used to miti­
models (e.g., support vector machine, classification and regression tree, gate flood risk through establishment of upstream check-dams to control
and mixture discriminant analysis) with MCDM models can assist local river flow and maintain the suspended sediment load and downstream
authorities for identifying susceptible areas for flood management dredging and cleaning activities of rivers and channels by construction
(Shikhteymour et al., 2023). Of the three models (gradient boosting of protective walls on the river bank.
decision tree (GBDT), XGBoost and CNN) used to map flood risk, the Our suggested methodology has a high capability to analyze large
GDBT model produced the most accurate results (Chen et al., 2021). datasets for spatial purposes in different scientific fields and to provide a
Overall, developing more accurate approaches for assessing flood risk basis for mapping floods over large spatial scales. Flood risk is depen­
and improving our understanding of flood hazard and vulnerability can dent on various features, and integrated modeling based on DL and
help lawmakers and decision-makers to mitigate damages and mortal­ MCDM models hence allows us to pin-point the most important of these
ities (Nasiri et al., 2016). for the prediction of flood risk. Additional accuracy of the predictive
models used to map flood risk may be achieved with higher spatial
resolution provided by remote sensing data, especially satellite images
3.5. Historical floods in the study area, and flood of January 2022 used to map vegetation cover and land use type, and the inherently
higher spatial resolution of DEM (e.g., <10 m) and soil maps.
The Hormozgan province and Minab have experienced floods for the
past several decades. For example, the Hormozgan province experienced 4. Conclusions
a big flood in 1992, and this flood had many negative economic and
social effects, and about 1000 casualties. Another flood that occurred in For the first time, this study developed a novel integrated modeling
January 2019 in Minab caused massive damage to the agricultural, approach based on an mLSTM deep learning model and an MCDM
natural resources, and urban sectors. ensemble model for assessing the spatial modelling of flood risk. The
The flood of January 2022 in Minab was initiated by heavy rainfall mLSTM model was used to map flood hazard, and a new MCDM
and hail associated with severe thunderstorms. Minab received more ensemble model (CODAS-EDAS-MOOSRA model) was developed based
than 270 mm of rainfall over two days, which filled the Esteghlal Dam on the Spearman’s rank correlation coefficients to map flood vulnera­
(located upstream of Minab city) to over 91% of its capacity. The same bility. Then, integration of the mLSTM and the CODAS-EDAS-MOOSRA
applied to the Shamil-Nian Dams. Inflow and outflow discharges to the ensemble model was used to map flood risk.
reservoir of the Esteghlal Dam were estimated at 13,000 m3/s and 2500 Based on the flood risk map, 27.4%, 34.4%, 14.8%, 15.7% and 7.7%
m3/s, respectively. Due to the opening of the Esteghlal Dam sluices to of the total area were classified as very low, low, moderate, high and
release water, a significant part of the downstream land was impacted by very high risk, respectively. The flood risk map provides good infor­
the resulting flood. The roads and water transmission pipes to the mation for the sustainability and comprehensive management of the
numerous villages and residential regions were severely damaged, and watershed, and the mitigation of detrimental flood effects. Generally,
significant parts of the crops were buried under flood waters and

9
A. Mohammadifar et al. Journal of Environmental Management 345 (2023) 118838

our novel methodology can be helpful for spatial assessments of various Costache, R., Tin, T.T., Arabameri, A., Crăciun, A., Ajin, R.S., Costache, I., et al., 2022a.
Flash-flood hazard using deep learning based on H2O R package and fuzzy-
natural hazards around the world. Overall, we recommend that future
multicriteria decision-making analysis. J. Hydrol. 609, 127747.
research should employ data with high spatial resolution (e.g., variables Costache, R., Arabameri, A., Costache, I., Crăciun, A., Islam, A.R.M.T., Abba, S.I., et al.,
extracted from DEM and satellite images) to improve the accuracy of 2022b. Flood susceptibility evaluation through deep learning optimizer ensembles
predictive models and the performance of other DL models (e.g., and GIS techniques. J. Environ. Manag. 316, 115316.
Darko, A.P., Liang, D., 2020. Some q-rung orthopair fuzzy Hamacher aggregation
gcForest, BiGRU, BiLSTM, GCN, etc.) to test flood risk mapping in key operators and their application to multiple attribute group decision making with
plains and watersheds in different climates. modified EDAS method. Eng. Appl. Artif. Intell. 87, 103259. .
Ekwueme, B.N., Agunwamba, J.C., 2021. Trend analysis and variability of air
temperature and rainfall in regional river basins. Civil Engineering Journal 7 (5),
Ethics approval and consent to participate 816–826.
Ekwueme, B.N., 2022. Machine learning based prediction of urban flood susceptibility
Not applicable. from selected rivers in a tropical catchment area. Civil Engineering Journal 8 (9),
1857.
Fernández, D.S., Lutz, M.A., 2010. Urban flood hazard zoning in Tucumán Province,
Consent to participate Argentina, using GIS and multicriteria decision analysis. Eng. Geol. 111 (1–4),
90–98. .
Gholami, H., Mohammadifar, A., Malakooti, H., Esmaeilpour, Y., Golzari, S.,
Not applicable. Mohammadi, F., et al., 2021a. Integrated modelling for mapping spatial sources of
dust in central Asia-An important dust source in the global atmospheric system.
Atmos. Pollut. Res. 12 (9), 101173.
Consent for publication
Gholami, H., Mohammadifar, A., Golzari, S., Kaskaoutis, D.G., Collins, A.L., 2021b. Using
the Boruta algorithm and deep learning models for mapping land susceptibility to
Not applicable. atmospheric dust emissions in Iran. Aeolian Research 50, 100682.
Gholami, H., Mohammadifar, A., 2022. Novel deep learning hybrid models (CNN-GRU
and DLDL-RF) for the susceptibility classification of dust sources in the Middle East:
Funding a global source. Sci. Rep. 12 (1), 19342.
Gholami, H., Mohammadifar, A., Fitzsimmons, K.E., Li, Y., Kaskaoutis, D.G., 2023.
Modeling land susceptibility to wind erosion hazards using LASSO regression and
Not applicable.
graph convolutional networks. Front. Environ. Sci. 11, 1187658.
Ghorabaee, M.K., Zavadskas, E.K., Olfat, L., Turskis, Z., 2015b. Multi-criteria inventory
CRediT authorship contribution statement classification using a new method of evaluation based on distance from average
solution (EDAS). Informatica 26 (3), 435–451. .
Ghorabaee, M.K., Amiri, M., Zavadskas, E.K., Hooshmand, R., Antuchevičienė, J., 2017.
Aliakbar Mohammadifar: Software, Formal analysis. Hamid Fuzzy extension of the CODAS method for multi-criteria market segment evaluation.
Gholami: Software, Formal analysis, Investigation, Visualization, J. Bus. Econ. Manag. 18 (1), 1–19. .
Glenn, E.P., Morino, K., Nagler, P.L., Murray, R.S., Pearlstein, S., Hultine, K.R., 2012.
Writing – original draft, Supervision, Project administration, Review &
Roles of saltcedar (Tamarix spp.) and capillary rise in salinizing a non-flooding
editing.. Shahram Golzari: Investigation, Visualization, Supervision. terrace on a flow-regulated desert river. J. Arid Environ. 79, 56–65. .
Hochreiter, S., Schmidhuber, J., 1997. Long short-term memory. Neural Comput. 9 (8),
1735–1780. .
Declaration of competing interest Jongman, B., Ward, P.J., Aerts, J.C., 2012. Global exposure to river and coastal flooding:
long term trends and changes. Global Environ. Change 22 (4), 823–835. .
Khosravi, K., Nohani, E., Maroufinia, E., Pourghasemi, H.R., 2016. A GIS-based flood
The authors declare no competing interests. susceptibility assessment and its mapping in Iran: a comparison between frequency
ratio and weights-of-evidence bivariate statistical models with multi-criteria
Data availability decision-making technique. Nat. Hazards 83 (2), 947–987. .
Krause, B., Lu, L., Murray, I., Renals, S., 2017. Multiplicative LSTM for sequence
modelling. .
Data will be made available on request. Kundzewicz, Z.W., 2002. Non-structural flood protection and sustainability. Water Int.
27 (1), 3–13. .
Luppichini, M., Barsanti, M., Giannecchini, R., Bini, M., 2022. Deep learning models to
Appendix A. Supplementary data predict flood events in fast-flowing watersheds. Sci. Total Environ. 813, 151885. .
Ma, Y., Zhang, Z., Li, D., Tang, M., 2021. Deflated reputation using multiplicative long
Supplementary data to this article can be found online at https://doi. short-term memory neural networks. Future Generat. Comput. Syst. 118, 198–207. .
Malczewski, J., 2006. GIS-based multicriteria decision analysis: a survey of the literature.
org/10.1016/j.jenvman.2023.118838.
Int. J. Geogr. Inf. Sci. 20 (7), 703–726. .
Masuya, A., 2014. Flood vulnerability and risk assessment with spatial multi-criteria
References evaluation. In: Dhaka Megacity. Springer, Dordrecht, pp. 177–202.
Mohammadifar, A., Gholami, H., Golzari, S., 2023. Stacking-and voting-based ensemble
deep learning models (SEDL and VEDL) and active learning (AL) for mapping land
Arabameri, A., Rezaei, K., Cerdà, A., Conoscenti, C., Kalantari, Z., 2019. A comparison of
subsidence. Environ. Sci. Pollut. Control Ser. 30 (10), 26580–26595.
statistical methods and multi-criteria decision making to map flood hazard
Moore, I.D., Grayson, R.B., 1991. Terrain-based catchment partitioning and runoff
susceptibility in Northern Iran. Sci. Total Environ. 660, 443–458.
prediction using vector elevation data. Water Resour. Res. 27 (6), 1177–1191. .
Arabameri, A., Saha, S., Chen, W., Roy, J., Pradhan, B., Bui, D.T., 2020. Flash flood
Narayanamoorthy, S., Annapoorani, V., Kang, D., Baleanu, D., Jeon, J., Kureethara, J.V.,
susceptibility modelling using functional tree and hybrid ensemble techniques.
Ramya, L., 2020. A novel assessment of bio-medical waste disposal methods using
J. Hydrol. 587, 125007.
integrating weighting approach and hesitant fuzzy MOOSRA. J. Clean. Prod. 275,
Arora, A., Arabameri, A., Pandey, M., Siddiqui, M.A., Shukla, U.K., Bui, D.T., et al., 2021.
122587. .
Optimization of state-of-the-art fuzzy-metaheuristic ANFIS-based machine learning
Nasiri, H., Mohd Yusof, M.J., Mohammad Ali, T.A., 2016. An overview to flood
models for flood susceptibility prediction mapping in the Middle Ganga Plain, India.
vulnerability assessment methods. Sustainable Water Resources Management 2 (3),
Sci. Total Environ. 750, 141565.
331–336. .
Athawale, V.M., Chakraborty, S., 2011. A comparative study on the ranking performance
Ngo, P.T.T., Pham, T.D., Hoang, N.D., Tran, D.A., Amiri, M., Le, T.T., et al., 2021. A new
of some multi-criteria decision-making methods for industrial robot selection. Int. J.
hybrid equilibrium optimized SysFor based geospatial data mining for tropical
Ind. Eng. Comput. 2 (4), 831–850. .
storm-induced flash flood susceptible mapping. J. Environ. Manag. 280, 111858. .
Aytaç Adalı, E., Tuş Işık, A., 2017. The multi-objective decision making methods based
Nilsson, R., Pena, J.M., Björkegren, J., Tegnér, J., 2007. Consistent feature selection for
on MULTIMOORA and MOOSRA for the laptop selection problem. Journal of
pattern recognition in polynomial time. J. Mach. Learn. Res. 8, 589–612. .
Industrial Engineering International 13 (2), 229–237. .
Norouzi, G., Taslimi, M., 2012. The impact of flood damages on production of Iran’s
Bengio, Y., Simard, P., Frasconi, P., 1994. Learning long-term dependencies with
Agricultural Sector. Middle East J. Sci. Res. 12, 921–926. .
gradient descent is difficult. IEEE Trans. Neural Network. 5 (2), 157–166. .
Panahi, M., Jaafari, A., Shirzadi, A., Shahabi, H., Rahmati, O., Omidvar, E., et al., 2021.
Chen, J., Huang, G., Chen, W., 2021. Towards better flood risk management: assessing
Deep learning neural networks for spatially explicit prediction of flash flood
flood risk and investigating the potential mechanism based on machine learning
probability. Geosci. Front. 12 (3), 101076.
models. J. Environ. Manag. 293, 112810. .
Parry, M.L., Canziani, O., Palutikof, J., Van der Linden, P., Hanson, C. (Eds.), 2007.
Costache, R., Popa, M.C., Bui, D.T., Diaconu, D.C., Ciubotaru, N., Minea, G., Pham, Q.B.,
Climate Change 2007-impacts, Adaptation and Vulnerability: Working Group II
2020. Spatial predicting of flood potential areas using novel hybridizations of fuzzy
Contribution to the Fourth Assessment Report of the IPCC, vol. 4. Cambridge
decision-making, bivariate statistics, and machine learning. J. Hydrol. 585, 124808.
University Press.
.

10
A. Mohammadifar et al. Journal of Environmental Management 345 (2023) 118838

Pham, B.T., Luu, C., Van Dao, D., Van Phong, T., Nguyen, H.D., Van Le, H., et al., 2021. Tan, S., He, X., 2021. A Concise method for feature selection via normalized frequencies.
Flood risk assessment using deep learning integrated with multi-criteria decision arXiv preprint arXiv:1609.07959.
analysis. Knowl. Base Syst. 219, 106899. . Tehrany, M.S., Pradhan, B., Jebur, M.N., 2015. Flood susceptibility analysis and its
Pradhan, B., 2010. Flood susceptible mapping and risk area delineation using logistic verification using a novel ensemble support vector machine and frequency ratio
regression, GIS and remote sensing. J. Spatial Hydrol. 9 (2), . method. Stoch. Environ. Res. Risk Assess. 29 (4), 1149–1165. .
Ramayanti, S., Nur, A.S., Syifa, M., Panahi, M., Achmad, A.R., Park, S., Lee, C.W., 2022. Tsakiris, G., Kordalis, N., Tsakiris, V., 2015. Flood double frequency analysis: 2D-
Performance comparison of two deep learning models for flood susceptibility map in Archimedean copulas vs bivariate probability distributions. Environmental Processes
Beira area, Mozambique. The Egyptian Journal of Remote Sensing and Space Science 2 (4), 705–716. .
25 (4), 1025–1036. Watershed Department of Iran (WDI), 2002. Statistics of Flooding Area in Iran Report,
Rao, R.V., 2007. London. Decision Making in the Manufacturing Environment: Using p. 50.
Graph Theory and Fuzzy Multiple Attribute Decision Making Methods, vol. 2. Winsemius, H.C., Van Beek, L.P.H., Jongman, B., Ward, P.J., Bouwman, A., 2013.
Springer, , p. 294. A framework for global river flood risk assessments. Hydrol. Earth Syst. Sci. 17 (5),
Rezaei, M., Mohammadifar, A., Gholami, H., Mina, M., Riksen, M.J., Ritsema, C., 2023. 1871–1892. .
Mapping of the wind erodible fraction of soil by bidirectional gated recurrent unit Wu, Z., Zhou, Y., Wang, H., Jiang, Z., 2020. Depth prediction of urban flood under
(BiGRU) and bidirectional recurrent neural network (BiRNN) deep learning models. different rainfall return periods based on deep learning and data warehouse. Sci.
Catena 223, 106953. Total Environ. 716, 137077. .
Scheuer, S., Haase, D., Meyer, V., 2011. Exploring multicriteria flood vulnerability by Yari, A., Ardalan, A., Ostadtaghizadeh, A., Zarezadeh, Y., Boubakran, M.S., Bidarpoor, F.,
integrating economic, social and ecological dimensions of flood risk and coping Rahimiforoushani, A., 2019. Underlying factors affecting death due to flood in Iran:
capacity: from a starting point view towards an end point view of vulnerability. Nat. a qualitative content analysis. Int. J. Disaster Risk Reduc. 40, 101258. .
Hazards 58 (2), 731–751. . Yesilnacar, E.K., 2005. The Application of Computational Intelligence to Landslide
Shikhteymour, S.R., Borji, M., Bagheri-Gavkosh, M., Azimi, E., Collins, T.W., 2023. Susceptibility Mapping in Turkey, vol. 200. University of Melbourne, Department. .
A novel approach for assessing flood risk with machine learning and multi-criteria
decision-making methods. Appl. Geogr. 158, 103035.

11

You might also like