Please download to get full document.

View again

of 17

Flood Susceptibility Assessment in Bangladesh Using Machine Learning and Multi-criteria Decision Analysis

This work proposes a new approach by integrating statistical, machine learning, and multi-criteria decision analysis, including artificial neural network (ANN), logistic regression (LR), frequency ratio (FR), and analytical hierarchy process (AHP).
2 views17 pages
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Documenttranscript
  Vol.:(0123456789)  1 3 Earth Systems and Environment https://doi.org/10.1007/s41748-019-00123-y ORIGINAL ARTICLE Flood Susceptibility Assessment in Bangladesh Using Machine Learning and Multi-criteria Decision Analysis Mahfuzur Rahman 1,2,3  · Chen Ningsheng 1  · Md Monirul Islam 3  · Ashraf Dewan 4  · Javed Iqbal 1,5  · Rana Muhammad Ali Washakh 1,2  · Tian Shufeng 1,2 Received: 29 July 2019 / Accepted: 21 September 2019 © King Abdulaziz University and Springer Nature Switzerland AG 2019 Abstract This work proposes a new approach by integrating statistical, machine learning, and multi-criteria decision analysis, includ-ing artificial neural network (ANN), logistic regression (LR), frequency ratio (FR), and analytical hierarchy process (AHP). Dependent (flood inventory) and independent variables (flood causative factors) were prepared using remote sensing data and the Mike-11 hydrological model and secondary data from different sources. The flood inventory map was randomly divided into training and testing datasets, where 334 flood locations (70%) were used for training and the remaining 141 locations (30%) were employed for testing. Using the area under the receiver operating curve (AUROC), predictive power of the model was tested. The results revealed that LR model had the highest success rate (81.60%) and prediction rate (86.80%), among others. Furthermore, different combinations of the models were evaluated for flood susceptibility mapping and the best combination ( 11 C) was used for generating a new flood hazard map for Bangladesh. The performance of the 11 C integrated models was also evaluated using the AUROC and found that integrated LR-FR model had the highest predictive power with an AUROC value of 88.10%. This study offers a new opportunity to the relevant authority for planning and designing flood control measures. Keywords  AHP · ANN · Bangladesh · Flood susceptibility map · FR · LR 1 Introduction Bangladesh is one of the most disaster-prone countries in the world. Flat topography, shallow riverbed, severe mon-soonal rainfall, and huge discharge of sediments are major factors responsible for floods in Bangladesh (Hossain 2015; Rahman et al. 2007; Sinha 2007). Therefore, identifying areas prone to floods is very important to reduce the loss of lives and properties. The flood event of 2017 (includ-ing floods in 1954, 1955, 1974, 1987, 1988, 1995, 1998, 2004, 2007, and 2014) caused enormous damage to prop-erty and considerable loss of lives. The heavy rainfall from *  Chen Ningsheng chennsh@imde.ac.cn Mahfuzur Rahman mfz.rahman@iubat.edu Md Monirul Islam mmislam@iubat.edu Ashraf Dewan a.dewan@curtin.edu.au Javed Iqbal  javediqbalgeo@gmail.com Rana Muhammad Ali Washakh washakh@qq.com Tian Shufeng tiansf@imde.ac.cn 1  Key Laboratory for Mountain Hazards and Earth Surface Process, Institute of Mountain Hazards and Environment, Chinese Academy of Sciences (CAS), Chengdu 610041, People’s Republic of China 2  University of Chinese Academy of Sciences, Beijing 100049, People’s Republic of China 3  Department of Civil Engineering, International University of Business Agriculture and Technology (IUBAT), Dhaka, Bangladesh 4  Spatial Sciences Discipline, School of Earth and Planetary Sciences, Curtin University, Kent St, Bentley, WA 6102, Australia 5  Department of Earth Sciences, Abbottabad University of Science and Technology, Abbottabad, Pakistan   M. Rahman et al.  1 3 upper-basin (upper and lower Brahmaputra, Kyaichinang, Barak, and so on) and lower-basin (the Ganges–Brahma-putra–Meghna basin) was accountable for severe flood-ing in Bangladesh in 2017. Because it turns into runoffs, due to the rough terrain and the vegetation situation in the area and such runoffs soon flow as floodwater. More than 30% areas were affected by the flood in 2017 (Uddin et al. 2019). It disrupted daily life, causing at least 134 deaths and nearly affected six million people across the country (Uddin et al. 2019). Cumulative discharge in the Brahmaputra and the Jamuna rivers within the country was increasing due to excessive rainfall in China, Nepal, and India, and as a result, water could not be drained out properly into the Bay of Bengal as reported by Bangladesh Water Development Board (BWDB).Flood susceptibility mapping can be defined as a quan-titative or qualitative assessment of the classification, area, and spatial distribution of flood, which exists or potentially may occur in an area. Therefore, flood susceptibility map-ping can help policymakers and relevant authorities to create emergency plans. It was stated that the occurrence of flood hazards cannot be stopped, but damages from flood could be avoided or substantially reduced if flood-affected areas were identified accurately (Sahoo and Sreeja 2015). Therefore, flood susceptibility assessment is very crucial for disasters alleviation. A broad range of model has been suggested by researchers to assess flood hazards. Most of the recent mod-els were mainly focused on hydrological models, hydrody-namic models, multi-criteria decision analysis (MCDA), statistical models (SM), and machine learning (ML) tech-niques incorporated into geographical information system (GIS) (Danumah et al. 2016; de Brito and Evers 2016; Elsafi 2014; Fernández and Lutz 2010; Lee et al. 2012; Luu et al. 2018; Rahmati et al. 2016c; Rao 2017; Shafapour Tehrany et al. 2017; Tehrany et al. 2014a; Yang et al. 2014). GIS and remote sensing are also important tools, which have been used extensively for hazard assessment (Ashley et al. 2014; Barua et al. 2016; Fernández and Lutz 2010; Islam and Sado 2000a; Kia et al. 2012; Luu et al. 2018; Shafapour Tehrany et al. 2017; Tehrany et al. 2014b). Studies have revealed that MCDA models are better for flood assessment. AHP is a popular model in the field of MCDA, because it can solve complex decision problem without any data (Danumah et al. 2016; Fernández and Lutz 2010; Luu et al. 2018). Besides, the most popular machine learning and statistical models in natural hazards analysis are artificial neural networks (ANN) (Elsafi 2014; Kia et al. 2012), logistic regression (LR) (Ara- bameri et al. 2018; Hong et al. 2015; Shafapour Tehrany et al. 2017; Tehrany et al. 2014a), frequency ratio (FR) (Pradhan and Lee 2010; Samanta et al. 2018b; Tehrany et al. 2019), weight-of-evidence (WoE) (Shafapour Tehrany et al. 2017; Tehrany et al. 2014b), and support vector machine (SVM) (Chen et al. 2018; Hong et al. 2015; Tehrany et al. 2015b). These models have perfect and consistent predic-tion capability for flood hazard occurrences (Bui et al. 2018; Chapi et al. 2017; Tehrany et al. 2014a), while hydrological and hydrodynamic models have some limitations, including time-consuming, requires careful and accurate calibration to yield accurate estimates of flood affected areas (Asare-Kyei et al. 2015; Fenicia et al. 2014). Although many researchers have conducted flood studies in various locations worldwide (Chapi et al. 2017; Dewan et al. 2007; Khosravi et al. 2016a; Masood and Takeuchi 2012; Seejata et al. 2018; Tingsanchali and Karim 2005), to the best of our knowledge, none of them integrated ML, SM models, and MCDA models for the development of flood susceptibility mapping, particularly for Bangladesh. Moreo-ver, the traditional method for flood susceptibility mapping in Bangladesh is the hydrological and hydrodynamic mod-els, which require input data and parameters from meteorol-ogy, river cross-sections, and discharge from both upstream and downstream (Khosravi et al. 2016b, 2018). These data are mostly unavailable for many areas, due to inadequate hydro-meteorological stations. At present, flood inundation area map is produced by the Flood Forecasting and Warn-ing Centre (FFWC) by comparing river water level with a coarse resolution (cf. 500 m) digital elevation model (DEM). Unfortunately, a high-resolution DEM and infrastructures data are not available (Bates 2004). Flood susceptibility assessment based on water level observation is not effective in providing spatially distributed flooding areas for timely monitoring of flooding event (Lin et al. 2019; Uddin et al. 2019). Therefore, the techniques used in the present study proved to be the best opportunity for relatively large and complex areas.The main objective of this research is to derive the extent of flood susceptibility areas in Bangladesh using four mod-els: artificial neural network (ANN), analytical hierarchy process (AHP), logistic regression (LR), and frequency ration (FR). The flood hazard map for Bangladesh was developed previously by considering flood frequency, flood duration with digital elevation data (Islam and Sado 2000b, 2002; Masood and Takeuchi 2012; Tingsanchali and Karim 2005), while in this study we have proposed nine causa-tive factors for flooding, i.e. rainfall, elevation, slope, flood depth, soil tract, geology, drainage area, flood duration, and land cover and land use (LULC). Besides, applying only one model will not be adequate to predict the susceptible areas in a study. Because, these models are mostly site specific and some research has confirmed that each model has its advantages and disadvantages. Therefore, the second objec-tive is to derive an integrated model, considering the best performing models, to develop a unique flood hazard map of Bangladesh, because model integration is expected to allow more precise assessment (Chapi et al. 2017; Costache and Zaharia 2017; Khosravi et al. 2016a; Mojaddadi et al. 2017;  Flood Susceptibility Assessment in Bangladesh Using Machine Learning and Multi-criteria…  1 3 Shafapour Tehrany et al. 2017). The key contributions of this research are: (i) to generate relevant models for the determi-nation of flood susceptible areas; and (ii) produce new flood hazard map for Bangladesh, using an integrated model. 2 Materials and Methods 2.1 Study Area The climatic condition makes Bangladesh the most vul-nerable country in the world to multiple hazards. This is a nation of over 162.7 million people (Bangladesh Bureau of Statistics B 2019) with a geographical area of 1, 47,570 sq.km located between 20º34 ′ N and 88º01 ′ E to 26º38 ′ N and 92º41 ′ E (Hasan et al. 2017). An annual growth rate of the population of Bangladesh is 1.37%, and therefore, is one of the most densely populated countries in the world hav-ing a population density of 1062.5 per sq.km (Hasan et al. 2017). A location map of Bangladesh is shown in Fig. 1. The country has 492 sub-districts and it is divided into five main physiographic regions, namely north Bengal region, northeastern region, Tippera-Comilla region, southwestern region, and Chittagong region, with various subdivisions (Islam and Sado 2000b). It is crisscrossed by three mighty rivers: the Ganges, Brahmaputra, and Meghna, popularly known as GBM. The alluvial soil deposited by these rivers has created highly fertile lands. It has three distinctive fea-tures: (i) a broad alluvial plain subject to frequent flooding, (ii) a slightly elevated relatively older plain, and (iii) a small hilly region drained by flashy rivers. The alluvial plain is a part of the larger plain of the Bengal, which is sometimes called the Lower Gangetic Plain. Elevations of the plains are less than 10 m above sea level. The hilly areas of the south-eastern region of Chittagong, the northeastern hills of Sylhet and highlands in the north and northwest are of low height. The Chittagong Hills constitute the only significant hilly system in the country. The climate of the country is tropi-cal and humid. The annual average rainfall varies between 2200 mm and 2500 mm, whereas extreme rainfall varies from 1200 mm to 6500 mm. The average temperature varies from 25 to 35 °C, during a year. 2.2 Data Preparation Rainfall data were collected from Bangladesh Water Devel-opment Board (BWDB) and NOAA satellite images gener-ated by NASA’s global precipitation measurement (GPM) mission. The digital elevation model (DEM) data with a spatial resolution of 300 m were obtained from Institute of Water Modeling (IWM) (Islam and Sado 2000b). The slope layer was extracted from DEM. A LULC map was obtained from the existing map produced by Space Research and Remote Sensing Organization (SPARRSO) (Islam and Sado 2000a) and updated by the Forest Department of Bang-ladesh in 2016 (Department 2016). The geological map of Bangladesh was obtained from Geological Survey of Bang-ladesh. The soil tract map with 1:100,000 scale for the study area was acquired from the Bangladesh Agricultural Research Council (BARC) and Soil Resource Development Institute (SRDI). The drainage areas’ data were collected from Bangladesh Agricultural Research Council (BARC). The flood depth in the study area was calculated by sub-tracting the land elevation from the computed flood water level (Tingsanchali and Karim 2005). The flood duration was determined using satellite images having spatial resolution of 12.5 m developed by International Centre for Integrated Mountain Development (ICIMOD) from Advanced Land Observing Satellite-2 (ALOS-2), Phased Array L-band Syn-thetic Aperture Radar (PALSAR) and Sentinel-1) of June 24,  July 17, August 15, and August 24 of 2017. The 2017 flood inundated most of the floodplain areas and lasted for more than 24 days according to field investigation done by flood forecasting department and observed hydrological data. 2.3 Computing Flood Inundation Area Flood inundation areas were calculated through remote sens-ing data analysis, Mike-11 hydrological model outputs, and three severe historical flood events of 1988, 1995, and 1998. Fig. 1 Location of the study area   M. Rahman et al.  1 3 The details of calculations are as follows: (i) first, we con-sidered four flood inundation maps of June 24, 2017, July 17, 2017, August 15, 2017, and August 24, 2017, which were prepared by ICIMOD considering the remote sensing imageries from ALOS-2, PALSAR, and Sentinel-1 imagery (ICIMOD 2017). The images were transformed into coordi-nate system (WGS 1984/UTM45N) based on an administra-tive map of Bangladesh. (ii) Second, Mike-11 hydrodynamic model output maps of Bangladesh Water Development Board were used to calculate flood inundation area consid-ering the same dates of June 24, 2017, July 17, 2017, August 15, 2017, and August 24, 2017. The Mike-11 model solves the unsteady free surface flow equations of continuity and momentum (Tingsanchali and Karim 2005). The DEM hav-ing 300 m spatial resolution was used with Mike-11. Finally, remote sensing imageries and Mike-11 model output maps were compared to estimate flood inundation areas. To cre-ate the final flood inventory map, National Oceanic and Atmospheric Administration (NOAA) advanced very high-resolution radiometer (AVHRR) data for the flood events of 1988, 1995, and 1998 were incorporated with flood inunda-tion areas (the inundated areas that did not appear in any of the images mentioned in this study were considered to be non-flooding areas, while the inundated areas appeared in all images were considered to be flooding areas), which were used to analyze the correlation between flood and flood conditioning factors. A flood inventory is a detailed register of the distribution and characteristics of past flood events. In the present study, the presence of flood was consigned a value of 1, while the absence of flood was consigned a value of 0 for preparing flood inventory map from flood inundation areas (Bui et al. 2018; Darabi et al. 2019). Finally, the values of all the flood conditioning factors were extracted to flood-ing and non-flooding points to form training and validation datasets. We identified 475 flood locations from where 70% of the locations were randomly selected for training and the rest were considered for validation. 2.4 Factors Affecting Flood Susceptibility To prepare flood susceptibility maps, various thematic lay-ers were used as conditioning factors. Rainfall is the main triggering factor that causes underground hydrostatic level and water pressure to increase. Mostly, heavy rainfall from upstream point (India) is the major reason for occurrence of flood in Bangladesh. The recorded rainfall amount dur-ing the monsoon period of 2017 varies from 120.10 mm to 898.62 mm (BWDB 2017). Land elevation is another factor in the assessment of flood susceptibility (Rizeei et al. 2019). Runoff flows from high to low lands, there-fore the probability of flood occurrence in low-elevated areas increases. Sometimes, lowland areas did not flood, while some high land elevation areas were flooded, due to flash floods in the northeastern part of the country. The elevation of the study area ranges from 0 to above 80 m of mean sea level. The likelihood of a flood increases, as the slope of a location decreases, making it a reliable indicator for flood susceptibility. Therefore, slope plays a major role in flooding and it also affects the direction of water flow. Land use and land cover map is one of the most important factors affecting floods, because vegetated areas are less susceptible for flooding due to the nega-tive correlation between a flooding event and vegetation density (Mojaddadi et al. 2017). However, urban areas are typically composed of impermeable surfaces and increased surface runoff, therefore it can be concluded that runoff conditions vary under different LULC patterns. Besides, LULC has a direct impact on a number of parameters in the hydrologic cycle, including interception, infiltration, and concentration, and therefore indirectly on flooding. Together, these characteristics yield information about the hydrological response and the degree of flood hazard (Islam and Sado 2000b). The geological factor of the study area is covered with various types of units, which directly or indirectly influence infiltration and runoff generation, depending on the porosity and permeability of soil and rock (Rahmati et al. 2016b). Moreover, geology signifi-cantly affects the formation of the drainage pattern that relates to the generation of floodplain (Bui et al. 2019). Islam and Sado (2000a) reported that largely impermeable surface geology areas are more susceptible to flooding. Therefore, geological units play an important role. The water infiltration initially depends upon soil properties (Rahmati et al. 2016b); therefore, soil group is another important conditioning factor. Todini et al. (2004) and Nyarko (2002) mentioned that the soil type plays a role in determining the water holding characteristics of an area, and hence affects flood susceptibility. Moreover, flood depth and flood duration directly contribute to flood occur-rence. The classification methods of flood conditioning factors are shown in Table 1 and thematic maps for flood conditioning factors are shown in Fig. 2.Furthermore, multicollinearity among all the factors was checked using the tolerance (TOL) and variance inflation factor (VIF), since linear collinearity between the conditioning factors will decrease the model prediction accuracy (Rahmati et al. 2016a). The coefficient values of TOL and VIF were below 0.10 and above 10.0, respec-tively, indicating the existence of collinearity between conditioning factors (Arabameri et al. 2019; Chen et al. 2018). The coefficient values for TOL < 0.10 and VIF > 10 indicated high multicollinearity between factors being considered. The results of the multicollinearity showed that no multicollinearity was present, among the nine fac-tors used (Table 2).  Flood Susceptibility Assessment in Bangladesh Using Machine Learning and Multi-criteria…  1 3 2.5 Modeling Approaches 2.5.1 Artificial Neural Network (ANN) ANN is a mathematical model of human perception that can be trained for performing a particular task on the basis of available dataset, especially to explore the relationship between inputs and outputs (Rauter and Winkler 2018; Valencia and Graña 2018). The most common type of ANN consists of three interconnected layers: (i) input layer, (ii) hidden layer, and (iii) output layer. The input layer receives data from different sources. The number of hidden layers and their neurons is often defined by trial and error (Elsafi 2014;  Jain et al. 1996; Karsoliya 2012). The number of neurons in output layers is fixed by the application and is represented by the class being processed. In this study, one of the most commonly used neural network methods, i.e.; multilayer perceptron (MLP) neural network was adopted (Kia et al. 2012). To apply the MLP neural network, the back propaga-tion (BP) algorithm with the sigmoid transfer functions was used in the hidden and output layers. Then, all observations were presented to the network, the weights were determined from the model by considering nine input layers, five hidden layers, and one output layer to produce a flood susceptibility map. The neurons in the input layer denoted different condi-tioning factors. The numbers of the hidden layers were con-firmed by running the MLP neural networks several times to gain compatible training and testing accuracies (Arora et al. 2004). The ANN model was trained with a maximum of 500 iterations and 10 tours with fivefold cross-validation. The convergence criterion was 0.00001. The probability of flood susceptibility (output layer) falls in the range between 0 and 1. When the percentage of the incorrect predictions in the neural network analysis decreased then the weights ( w i ) were stored to calculate flood susceptibility scores (FS). The Gra-dient Descent was used to estimate weights, where the initial learning rate, lower level learning rate, and the momentum were 0.4, 0.001, and 0.9, respectively. Moreover, the inter-val centre and interval offset were 0 and ± 0.5, respectively. Basically, the weights were calculated by normalizing and Table 1 Flood conditioning factors and classification scheme for flood susceptibility assessmentFlood conditioning factorsClassesMethod and referenceRainfall (mm)(i) Above 600; (ii) 401 to 600; (iii) 201 to 400; and (iv) 0 to 200.Equal interval [(Pham et al. 2017)]Elevation (m)(i) 0 to 4; (ii) 4 to 8; (iii) 8 to 12; (iv) 12 to 16; (v) 16 to 20; (vi) 20 to 40; (vii) 40 to 60; (viii) 60 to 80; and (ix) above 80.Manual [(Islam and Sado 2000a)]Slope ( ° )(i) 0 °  to 10 ° ; (ii) 10 °  to 20 ° ; (iii) 20 °  to 30 ° ; (iv) 30 °  to 40 ° ; (v) 40 °  to 50 ° ; (vi) 50 °  to 60 ° ; (vii) 60 °  to 70 ° ; and (viii) 70 °  to 80 ° .Equal interval [(Ouma and Tateishi 2014; Rahmati et al. 2016c; Seejata et al. 2018)] LULC(i) Cultivated land; (ii) boro rice field; (iii) cultivated lowland; (iv) dry fallows; (v) mixed cropped areas; (vi) mangrove area; (vi) highland with mixed forest; (viii) highland with settle-ments; (ix) saline area; and (x) watercourse/river.Supervised classification [(Islam and Sado 2000a)]Geology(i) Coastal deposits: beach and dune sand; (ii) deltaic deposits: silt, sand, tidal mud, and so on; (iii) alluvial deposits: alluvial sand, silt, clay, Chandina alluvium, valley alluvium, and so on; (iv) alluvial fan deposits: gravelly sand; (v) residual deposits; (vi) bedrock: Pleistocene and Pliocene; (vii) Tipam group: Pleistocene, Neogene, Tipam sand and stone; (viii) Surma group: Neogene, Miocene, and Oligocene; and (ix) major river.Supervised classification [(Islam and Sado 2000a)]Soil tract(i) River and water body; (ii) hill tract: red clay soil, fine sand, and the mixture of two; (iii) Barind Tract: deep reddish brown terrace soils, gray, and silty and poorly drained; (iv) coastal saline tract: saline and alkaline; (v) Madhupur tract or red soil tract: well to moderately well-drained, reddish brown to yellow–brown, strongly to extremely acidic, friable clay soils over deeply weathered, red-mottled, and Madhupur clay; (vi) Gangetic alluvial: clay loam, sandy loam, calcareous and non acidic; (vii) Tista silt: weekly acidic; and (viii) Brahma-putra alluvial: loamy soil.Supervised classification [(Islam et al. 2017; Shafapour Tehrany et al. 2017)]Drainage area (%)(i) 0 to 25; (ii) 25 to 50; (iii) 50 to 75; and (iv) Above 75.ManualFlood depth (m)(i) 0 to 0.50 m; (ii) 0.51 to 1.00 m; (iii) 1.01 to 1.50 m; (iv) 1.51 to 2.00 m; and (v) 2.01 to 2.50 m.Manual [(Tingsanchali and Karim 2005)]Flood duration(i) Very long; (ii) Long; (iii) Medium; (iv) Short; and (v) No flooding.Manual [(Islam and Sado 2000a)]
Advertisement
MostRelated
View more
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks
SAVE OUR EARTH

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...

Sign Now!

We are very appreciated for your Prompt Action!

x