Scientometric analysis of flood forecasting for Asia region and discussion on machine learning methods

1.1 Research significance

A significant study has been done in recent decades to investigate the factors that influence flood analysis with respect to climate change, and some valuable results have been reached. In the contemporary area, other assessment studies are carried out, although these are mostly manual analyses. This research examines the application of ML algorithms for flood forecasting in locations where indicators or advanced sensing devices are around. This article attempted to provide a complete overview of prior research on vulnerability, flood risks, and flood vulnerability. ML methods, based on past data, statistically describe the nonlinearity of the flood without having any knowledge of the basic physical methods and high efficiency related to physical systems with low processing costs, quick training, validation, testing, evaluation, and minimal intricacy. Furthermore, ML methods such as decision trees (DTs) accurately identify similar groups with varying susceptibility levels. This method works quickly and efficiently, and ANNs find long-term evaporation and transpiration and rainfall forecasting and present excellent abilities of ANN models. The wavelet neural networks (WNNs) develop a reliable streamflow prediction model that would be helpful in flood control efforts.

This work focused on developing countries and their strategies to cope with floods. Furthermore, the safeguards in place to alleviate the consequences of flash floods are insufficient. The study explored numerous forms of vulnerability and various vulnerability indicators and procedures for calculating the vulnerability index. The purpose of this article is to provide current flood analysis techniques, emphasizing the methodology employed. Highlighting current research goals and the essential procedures may assist in swiftly selecting the best research method and identifying critical gaps that should be addressed in future studies. This research study is divided into four sections. Section 1 introduces ML and the hydrogeological hazard as flood. Section 2 discusses the importance of research. Further, the methodology of Scientometric analysis is described in Section 3. Section 4 describes flood-prone areas, Section 5 presents the methods of ML used for flood forecasting, Section 6 is discussion, and Section 7 is about the conclusion and future results.

2.1 Publications per year

Each searched keyword’s yearly publishing trend has been presented. The term “Floods in Asia” yielded the majority of documents (3,374). Between 2004 and 2009, there were 1,132 total publications (sum of all keyword results), which might be considered a peak era. Following that, from 2010 to 2019, there was a significant increase in the number of publications (total: 1,744), but a steady reduction in comparison to the preceding 6 years, and this era may be characterized as a pick-up and pace phase. It is amazing to observe how academics are increasingly emphasizing environmentally friendly study in their studies. It is crucial to note that while searching for various keywords, there is the possibility of finding duplicate articles, which are then added to the total number of publications, which is a restriction of our research. Figure 2 shows the number of publications per year.

Figure 2

Number of publications per year.

2.2 Journal publishing

Source mapping allows for the evaluation of system analysis and design visualization. These resources provide us with the access to information that is limited by certain restrictions. The research pattern may be used progressively in the analysis area after initializing the mapping of research sources. This research was carried out using the VOS viewer and Scopus bibliometric data. The “analysis type” was selected as “bibliographic coupling,” and the “analysis unit” was chosen as “sources.” The minimum number of documents required for a source was set at 20, and 13 of the 605 sources satisfied this requirement. Table 1 shows the top sources/journals that published at least 50 papers on waste material usage in concrete, as well as the number of citations and overall link strength. Flood analysis, journal of hydrology, and hydrological processes, geomorphology, natural hazards, journal of geological society, and geophysical research are the top six journals in terms of published documents, contributing 22, 20, 15, 15, 10, and 9%, respectively. The top three journals in terms of citation count were the journal of hydrology, geomorphology, and hydrological process research, with 5,575, 2,668, and 2,582 citations, respectively (Figure 3).

Figure 3

Source mapping.

2.3 Keywords mapping

Keywords are useful research material that identify and reflect the study domain’s basic topic. Table 2 shows the top 12 terms in the current study area that appear most often in research publications. Asia, flooding with floods, flooding, Asia, China, and Eurasia are the most regularly used terms in the related subject, accounting for the top five. The keywords network co-occurrence visualization and linkages are shown in Figure 4. The frequency of a term is represented by the size of the keyword node, while the location of the keyword is represented by its co-occurrence in published articles. The picture also shows that the aforementioned keywords have more nodes than the others, meaning that they are the most important terms in the current research topic. Various colors have been used to indicate keyword clusters in the figure, suggesting keyword co-occurrence in different published articles. The colors red, green, and blue show three groups of keywords. This discovery will be valuable to authors in the future when selecting keywords and obtaining relevant articles.

Figure 4

Keywords mapping.

2.4 Top countries

Some countries have contributed more in the past and continue to do so in the current research area than others. Table 3 shows the major contributing countries in terms of the number of publications and citations in the subject field. In terms of total papers, China, the United States, and India were the top three contributors, with 543, 398, and 285 documents, respectively. The United States, China, and the United Kingdom were the top three participating nations in terms of citation count, with 23,383, 31,302, and 14,133 citations, respectively. The impact of a country on the development of the current research field is measured by the number of documents, citations, and overall link strength. The overall connection strength shows how a region’s publications affect other locations involved in these study fields. China has the greatest overall link strength (837), followed by the United States (789) and India (420). As a result, the nations in the aforementioned areas were deemed to be the most influential in the subject of sustainable development. The nations’ relationship in terms of citations is seen in Figure 5. The size of the frame depicts the region’s involvement in the relevant research area. The graphical depiction of participating nations will aid future academics in developing scientific relationships, generating joint venture reports, and exchanging innovative strategies and methodologies.

Figure 5

Top countries.

3.1 China

Floods are a commonly found hazard in China. Between 1950 and 2018, floods in China caused the death of 282,737 individuals, destroyed 6 billion acres of land, and economically cost $6,000 billion [54]. However, the external effects of storms were not addressed. Since 1950, the amount of flood-damaged land has increased, indicating a possible increase in flood damage, despite a decrease in flood-related death [55]. China is now one of the world’s largest most flood-prone nations, and its weather is anticipated to vary as much as the worldwide average this century [56]. In China, demographic vulnerability and possible damage from coastal flooding in the twenty-first century [56,57]. In contrast, using a modeled flood damage map, gridded land for development, and differentiated community datasets, spatiotemporal variations in urban land, and population in floodplains were explored, and fast increases in flood vulnerability were shown to be associated with urban growth [58]. Satellite-based nighttime light data, which provide a rare chance to actually monitor movement of people from space, have been extensively used as a proxy for human activities in the past few decades, thanks to the advent of numerous nighttime sensing devices [59]. At a given location, outburst floods have the ability to deliver substantially greater peak flows than climatic floods. Despite the fact that we catalogued 287 landslide dams, the majority of which had undergone outburst floods, numerical for just a handful of the floods was accessible [60]. Only 24 avalanche dam outbursts have peak flood flow values, 20 of which were explicitly documented and 4 via hydro-geological or uses the process data such as dammed reservoir layers, dam proportions, and flooding cross-section [61]. Flood exposure reduces from the eastern coastal regions to the middle and western regions in general. In Tianjin, Shanghai, Henan, Jiangsu, Shandong, and Beijing [62], the overall population vulnerable to floods is around 500 people per square kilometer. Shanghai has the most exposure community, with 2,556 people per square kilometer [63]. Major events of flood in China are shown in Table 4. Monsoon season of China is shown in Figure 6.

Figure 6

Rainfall season of China [66].

3.2 India

The Uttarakhand floods of June 16–17, 2013, wreaked havoc on lives and properties. Heavy rains in the Kedarnath valley caused flash floods in the Mandakini and Saraswati Rivers [67]. In addition, owing to glacier/melting snow and glacier mountain ranges, the Mahatma Gandhi Sagar glacial lake outburst flood (GLOF) enhanced runoff towards Kedarnath and downstream displays the region’s strong and isolated rainfall from June 14 to June 19, 2013 [68].

This resulted in an occluded discontinuity and pulsed extension of monsoon across the Himalayas during the Uttarakhand flood [69]. In the high troposphere, the chilly gradient of the frontal WD (warmer in the lead, cooler in the tail) and the warmer, more humid monsoon flow in the lower troposphere caused this discontinuity [70]. Consequently, a thorough knowledge of these interactions was gained by high-resolution modeling and observational analyses. Floods occurred in Jammu and Kashmir and neighboring states as a result of depressions over Rajasthan recurving northward or northeastward, causing flash flooding [71]. Low pressure systems in the lower troposphere were coupled with a trough in the upper-level westerlies, which had an embedded jet stream 35, which resulted in a significant amount of moisture being absorbed [72]. The interplay between the low westward-moving monsoon and the eastward-moving deep trough in the mid-latitude westerlies 36 results in continuous heavy rainfall [73]. The state’s historical rainfall data indicates that the occurrence was unusual and may be classified as belonging to the “never before” group. The rainfall totals in Anantnag, Kukernag, and Quazigand topped prior records by 24, 48, and 72 h, respectively, while the rainfall totals at Katra beat previous records by 48 and 72 h [74].

The total rainfall reported in the Chennai district from November 1 to December 5, 2015 was 1416.8 mm, compared to the typical rainfall of 408.4 mm during the same period in 2014 [75]. In the Indian area, an anomalous mid tropospheric high located to the west of the region might cause anomalous northerly winds to blow to the north, preventing propagating storms from moving farther north and west [76]. Because of this, dry weather prevailed in the northwest Indian area while wet conditions prevailed in the southeast Indian peninsula; for example, an anomalous high pressure system over the Middle East around November resulted in higher rainfall across southern India [77].

While the rest of the world’s sea surface temperature has increased significantly over the last 35 years, temperatures in the northern and western Bay of Bengal have hardly changed at all. At this point, it is impossible to determine if global warming had a role in the incidence of excessive 1 day rainfall that caused extensive flooding in Chennai in December 2015 [78].

On December 1, 2015, heavy rainfall was caused by the creation of deep convective clouds that moved from the Bay of Bengal to Chennai city [79]. As a result of the obstruction of the natural drainage network, flood water remained stagnant in local depressions long after the water level in the Jhelum River had fallen. A number of studies conducted prior to the flood event have indicated that rapid urbanization, deforestation, and unrestrained land-use changes, such as encroachment in wetlands such as the littorals of the Dal and Anchar Rivers, Hokrasar Lake, and Narkara Lake, pose a significant threat to the water supply. Since 1972, the built-up area within Srinagar city limits has increased by approximately 29.20% (from 18.10– 84.50 km²), which is nearly three times the rate of population growth and observations clearly point to the encroachment of natural flood ways in the city [80]. These studies clearly stated that the depletion and degradation of wetlands as a result of excessive siltation, rapid urbanization, and encroachments have a negative impact on their ability to retain floodwaters, including flash flood waters during peak discharge, and that they require immediate restoration to function properly [77]. To effectively combat this type of natural hazard phenomenon, FS mapping is required. Based on these considerations, we investigated whether FS mapping may improve forecast accuracy in the Koiya River basin (Eastern India) [81]. Figure 7 shows the flood in eastern India.

Figure 7

Flood in Eastern India: (a) bridge under water, (b) flooding in large area, (c) car and houses flooded with water, and (d) local boat is the main medium of transportation during flooding [81].

3.3 Nepal

The Himalayas are the source of the most of Nepal’s long-lasting rivers [82]. Stable rainfall (snow) and mountains play a critical role within Nepal’s key water resources, while rainfall water is immediately released into rivers in a relatively short time frame, resulting in flash floods [83]. In Nepal, the monsoon season (June–September) accounts for around 80% of total yearly precipitation, whereas winter snowfall accounts for about 3% of total annual heavy rainfall [84]. During the dry season of the year, the rainwater that melts contributes greatly to the mountain regions’ continual base flow. The GLOF, on the other hand, is one of the major worries. Most of the glacier lakes in Nepal Himalayas have been embanked by a sediment dam built during the modernity era [85]. These dams are often unconsolidated and inherently unstable. The catastrophic flood might cause issues not just for the infrastructure but also for the people living downstream. As a result, mitigating methods like regulated breaching, syphoning surplus lake water, and constructing a route under the moraine barrier are required to manage the GLOF. Nepal has around 6,000 rivers and rivulets flowing through it. Snow-fed rivers like the Narayani, Karnali, and Koshi are perennial among them [86]. These rivers travel across the Terai plains and originate in the Himalayas. During the monsoon season, from June to September, the flow in rivers increase much more, affecting the adjacent settlements. The following years saw large floods in Nepal: Tinao Basin (1978), Koshi River (1980), Tadi River Basin (1985), Sunkoshi Basin (1987), and the devastating floods triggered by a cloud burst in the Kulekhani region (1993), which killed over 1,336 people [87]. The most recent Koshi flood, which occurred in 2008, harmed around 200,000 people. Between April and May of this year, an interdisciplinary field research of past GLOFs in Nepal’s Kanchenjunga area was done. Since 1921, at least six large GLOFs have occurred in the area, according to oral history and field data [88]. A remote sensing investigation validated the existence of the six GLOFs indicated by informants, as well as two minor flood events that happened before 1962 but were not documented. According to a computer simulation of the Nangama GLOF, it was initiated by an ice/debris avalanche releasing 800,000 m³ of material, generating a rush wave that crossed the final moraine and unleashed 11.2 × 10⁶ m³ ± 1.4 × 10⁶ m³ of water (Figure 8). Several sources in Tapetok stated that the 1963 Olangchun Gola 1 flood was worst they had ever seen, almost affecting their homes 100 m well above river [89] (Figure 9).

Figure 8

Nangama glacial lake and Chheche Pokhari 1980 GLOF [87].

Figure 9

Tapetok (1,380 m) on the Tamor. Flooding from the 1963 Olangchun Gola 1 GLOF almost reached the residences in the backdrop, or around 100 m above the river channel, according to eyewitnesses [87].

3.4 Pakistan

When it comes to floods, topography has a significant impact. There are two major mountain ranges in Pakistan, the Hindu Kush and Karakoram mountains. Between 1,000 and 8,000 m above sea level, the snow-clad summits of the northern parts may be found. If you are looking for an example of the world’s highest peaks (above 8,000 m), go no farther than K-2, Gasherbrum II, the Broad Peak and the Gasherbrum [90]. There are almost 68 mountains in Pakistan that are above 7,000 m high. Floods caused by heavy rain and cyclones are the most common. In addition, tiny dams have been breached, resulting in very high floods. A large portion of the nation (59.3%) gets rainfall of less than 200 mm per year. However, the Himalayas get between 760 and 1,270 mm of annual rainfall, which accounts for over two-thirds of the average annual rainfall in the Indus River System. More than 594,700 km² of land were inundated, affecting 166,075 communities, resulting in nearly $30 billion in direct damages and the deaths of 10,668 people over the course of the last 60 years [91]. During the rainy season, rain is one of the primary causes of flooding in Pakistan, which is often aided by river snowmelt. Snowmelt upstream of Tarbela Dam contributes to the rainy discharge in northern Pakistan. However, during the monsoon season, urban floods are widespread in Pakistan’s several cities, including Karachi, Islamabad, Lahore, and Hyderabad. Recent instances include the cyclones Yemyin and Phet that hit Pakistan in 2007 and 2010, respectively [75]. The 2010 floods caused unprecedented devastation and claimed the lives of about 1,985 people. Every province in Pakistan has a different kind of flooding. Consequently, it is a difficult challenge to solve, especially for those in charge of flood control. Many rivers have distinct catchment regions due to a variety of factors including geography, demographics, and socioeconomic situations. Climate change caused by human activity is cited by the Intergovernmental Panel on Climate Change as one of the causes driving up Asian monsoon rainfall levels. However, this condition may be difficult to meet until more scientific research is done. Flood embankments on the Indus River’s crest make the riverbed rise over the surrounding regions. As a result, the cheap regions remain wet for a longer length of time since the flow that exits the river does not rejoin the stream. Because Sindh’s tail links to the Arabian Sea, it is impossible to implement flood preventive actions in the province’s upper portion. From Guddu until a few kilometers short of the Arabian Sea, a double line of flood berms has been constructed on both banks of the river in the majority of the cracks. In Punjab, defensive marginal bunds have been constructed to maintain irrigation infrastructure and headworks, or to defend particular towns and communities. In the event that flood levels rise over the planned level, pre-determined breaching portions will be available [92]. In order to prevent erosion, spurs have been installed in high-risk regions. Most rivers in Balochistan have high bed slopes because of the region’s terrain. Besides the low-lying regions along the river banks, flash floods induce banks to crumble due to their force and velocity. There are embankment barriers or short spurs protecting agricultural land from the direct effect of flash floods [93]. Flood flow channels were built to disperse the surge of flood water. Irrigation water is also diverted from these sources. Most rivers in centrally governed regions have been plagued by bank erosion because of their high bed slopes. As a result, water bypasses have been built to spread the flood flow, which is also utilized for irrigation of agricultural community. The goal of this research is to determine the effect of the 2010–2011 flash floods on agricultural output in Sanghar, Sindh’s eastern district as shown in Figure 10. Agriculture is the major source of income for the over two million people that live there. Cotton, sugarcane, and wheat are the district’s principal cash crops. Sanghar was devastated by flash floods in 2010–2011 as a result of severe rains. These floods had a negative impact on the crops that were growing at the time. They did, on the other hand, improve the fertility of agricultural soils. Floods were quantified by mapping pre- and post-flood plant cover using satellite photos, precipitation data, and geographic information system (GIS) tools. From 2009 to 2013, supervised classification was used to derive vegetation area from Landsat photos [94]. The post-flood years show a large increase in plant cover, according to the temporal study. Furthermore, the research findings were corroborated by a comparison of pre- and post-flood agricultural production data obtained from the Pakistan Bureau of Statistics, which revealed an increase in crop output [95].

Figure 10

Sanghar distract [95].

3.5 Bangladesh

Bangladesh bears the repercussions of its placement in the Ganges–Brahmaputra–Meghna (GBM) river basin low-lying fluvial floodplain every year. Approximately 80% of Bangladesh’s territory is covered by flood plains, with up to 34% of the country’s land area submerged for 5–7 months each year [96]. Flooding is becoming more common in Bangladesh, according to research, and flooding is anticipated to become more common as a result of climate change. Flooding is a natural occurrence that occurs annually in Bangladesh. Bangladesh is susceptible nations to natural disasters due to its geography. Floods are one of the key obstacles to Bangladesh economic growth and planning; in other words, catastrophic floods, which occur regularly, are one of the country’s most critical handicaps [97]. Bangladesh is the second most flood-affected nation in the world, with a population of over a billion people. Flooding normally begins in the first week of June, lasts until October or November, and inundates nearly one-third of Bangladesh’s land area. This yearly flood phenomenon is typically beneficial to farmers. It is known as “Borsha,” and it assists farmers in cultivating their agricultural areas amid flood waters, as well as the “Aman” crops, which need flood water to develop correctly [98]. As a result, flooding is a helpful tool for watering and enriching agricultural regions. Even in years with regular rainfall, the monsoon populates 20–30% of the land each year. The regular series of floods in Bangladesh begins in April and May with severe flooding in the north and east highlands; in a flash flood, rivers rise fast and recede quickly, generally in a matter of days or hours. The enormous GBM River Basin receives most of the flood water. The monsoon has an impact on the whole GBM River. Heavy seasonal rainfall, which falls 80% of the time between June and October, and snow melt water from the Himalayas, find an outlet to the sea via the Bengal basin, which accounts for around 7.5% of the total river catchment of the main river systems. Flooding in Bangladesh has been caused by the building of barrages and protection works along the banks of the rivers, notably upstream (India, Nepal, and Bhutan). As a result of Bangladesh’s storm surge measures, such as the Brahmaputra right bank dam and Chalan beel dam barrier, floods have become more intense in areas that are not covered [99]. Precipitation of river-borne sediments is the fundamental process for delta formation. The projected embankments would prevent sedimentation on Bangladesh’s delta plain from aggravating riverbeds and causing the lands behind the embankments to be submerged by the increasing sea. Other issues that might arise as a result of the execution of such a megaproject include flash floods, significant river channel modifications, the loss of a large amount of land to the project, loss of land fertility, and possibly population displacement. Building such embankments would need ongoing maintenance for decades, which will be a difficult challenge for Bangladesh. Flood is one of Bangladesh’s most prevalent natural catastrophes. Bangladesh has been hit by various sorts of disasters practically every year in the past. Each flood has a severe impact on the country’s people, livelihoods, and properties. However, both the state and non-government organizations in Bangladesh have made a variety of steps to mitigate the negative effects of floods. According to this report, implementing a comprehensive flood management strategy and a coordinated strategy among the various flood control players can help Bangladesh reduce flood risk. The communities of Nowapara and Pashurbunia were chosen due to their proximity to the Bay of Bengal, socioeconomic instability (i.e., living below the poverty line), and susceptibility to cyclone catastrophes. These two groups are illustrative of the overall situation of coastal Bangladeshi populations. The Bangladesh Water Development Board owns the wall, which is situated between 20 and 500 m of the Rabnabad Channel and feeds both villages (Figure 11a and b). Majority of the houses are in close proximity to the canal and wall (Figure 11c and d).

Figure 11

(a and b) The community’s embankment and (c and d) houses positioned close to the Rabnabad Channel [99].

3.6 Sri Lanka

Examined the qualitative evaluation of flood risk in Sri Lanka western province. We employed an index that has only weighted variables for this investigation [100]. Risk, susceptibility, and vulnerability indexes are used to create this risk score. Thus, a risk, exposure, and relative risk maps have been created. Only one of the eight (0.9%) GN Divisions has a very high flood danger, while only four of the eight have a high flash flood. A flood risk map may be used to identify the GN Divisions at risk of flooding and determine the relevant disaster mitigation strategies before undertaking planned projects in the region. The resulting data may be used to influence choices about disaster preparation, early warning, and other measures that can be put in place to improve a vulnerable population’s adaptive ability and lower their catastrophe risk. The study’s findings also include recommendations regarding how much money should be set aside for disaster relief and recovery efforts once flood has occurred. Gumblel model was used for recovery process of Ratnapura Divisional Secretariat region. The major approach utilized in this research was a GIS, which included a digital terrain model and data from the previous 21 years of flood heights [101]. Low, medium, and high flood danger maps, as well as likelihood of inundation maps, were generated during the course of the research for 5, 10, 20, 30, 40, 50, and 100 years. The flood maps revealed that 64.3% of the study region is in the high-hazard zone. Ratnapura town, Ratnapura town west, Godigamuwa, Thiriwanaketiya, and Weralupe are among them. These flood maps are essential for flood disaster management, rescue and restoration programs, and more. In the Rathnapura area, flood risk reduction measures in vernacular dwellings were investigated. 15 case studies were used to gather information. Chronological analysis approaches were used to examine the case studies. To design flood risk reduction methods for the research region, five factors were established. Location and orientation, plan layout, substructure superstructure, and services are the criteria [102]. The lack of contemporary materials and professional expertise in the district’s case study region was discovered in this investigation. As a result, this research will be utilized to create materials and building procedures in flood-prone regions of Sri Lanka. The potential of a fifteenth century canal system in Sri Lanka’s Metro Colombo region to regulate floods was investigated. MIKE FLOOD modeled the canal system for 10, 25, and 50 year rainfall return periods to attain the desired flood levels. The flood levels, inundation distributions, and impacted populations were all taken into account while analyzing the effects of the rainfalls in question. In the research, two simulation scenarios based on river boundary conditions were run, and they were classed as favorable and least favorable. It was discovered that, under current circumstances, the canal system could only manage a 10 year rainfall flood event in the best-case scenario. As a result, the canal system’s long-term viability in the face of future catastrophic occurrences is doubtful. Four interventions were developed to reduce such floods, and their effects were assessed. The flood water levels were dropped locally when the countermeasures were implemented one at a time, and they were not up to the flood safety standards of the surrounding region. When all four countermeasures were implemented simultaneously, flood water levels were dramatically reduced below flood safety limits under a 50 year design rainfall under ideal conditions. When all four countermeasures were used in a combined manner, the flooded area was significantly reduced. In that situation, the flooding area was reduced by 46%, and the number of persons impacted was reduced by 49% [103].

4.1 DT algorithms

It is one of the contributors to predictive modeling, and it has a broad range of applications in flood simulation, including the ML technique of DT. DT makes use of a decision tree that extends from the branches to the goal values of the leaves. Classes are represented by leaves in classification trees, and features are represented by conjunctions of feature labels in DTs as shown in Figure 12. The final variables in a DT contain a discrete set of values, where leaves represent class labels and branches represent conjunctions of feature labels. An ensemble of trees is used to construct a regression tree when the target variable in a DT has continuous values and an ensemble of trees is used to construct the DT (RT) [104,105]. There are several parallels and distinctions between regression and classification trees. Due to the fact that DTs are considered fast algorithms, they have become particularly popular in ensemble forms for modeling and forecasting floods. The classification and regression trees are two types of trees (CART) [106,107]. The DT is a ML technique that may be used for classification and regression problems. It worked in a similar manner to SVM. The DT and the bagged decision are the same thing. The main difference is that the forecasting model does not use the whole set of data as input, as opposed to the previous method [108]. A variety of tree predictors are included in RF. A collection of response predictor values is generated for each individual tree, which is then connected with a set of independent values. Furthermore, an ensemble of these trees finds the most appropriate set of classes for a given situation [109]. The flood prediction modeling field, RF, has been proposed as a viable alternative to SVM, which in many cases provides superior performance [110]. It was determined that the performance of ANN, SVM, and RF in general applications to floods was the best, with RF delivering the greatest results [111]. In order to reduce the variance of the final variable, M5 divides the decision space into single characteristics and generates a DT from those single qualities. Additional decision-tree algorithms that are commonly used in flood prediction include reduced-error pruning trees, Naive Bayes trees, chi-squared automatic interaction detectors (CHAIDs), logistic model trees, alternating decision trees, and exhaustive CHAIDs, to name a few [112].

Figure 12

Classification diagram of DT [113].

4.2 ANN

ANN model is made up of input, hidden, and output layers and is built on the perceptron method, which is similar to human neurons. Every layer is made up of nodes, which are the fundamental computational units of a neural network [114]. Each node is linked to the others to use a feed forward mechanism that takes linkages into account while computing node simulation. A feed forward back propagation neural network with one input layer, one hidden layer, and one output layer is shown in Figure 13. As effective mathematical modeling systems with efficient parallel processing, ANNs are capable of simulating the behavior of a biological neural network by employing interconnected neuron units. ANNs are the most common learning algorithms among all ML techniques, and they are well-known for being adaptable and efficient at modeling complex flood processes with a high fault tolerance and precise approximation [115]. When compared to standard statistical models, the ANN technique was shown to be more accurate in terms of prediction accuracy [116]. ANNs, rather than relying on a catchment’s physical qualities, gain significance from past data. As a result, ANNs are regarded as trustworthy data-driven methods for developing black-box models of complicated and nonlinear rainfall and flood connections [117]. Though ANNs have several benefits, they also have a number of disadvantages when used in flood modeling, including network construction, data management, and physical interpretation of the studied system. When employing ANNs, the poor precision, the need to repeat parameter adjustment, and the delayed reaction time to gradient-based learning processes are all significant disadvantages to be aware of [118]. Modeled flash floods, using ANN and ANFIS, used a dataset compiled from 14 distinct streamflow gauge sites to get the desired results. The assessment criteria were the root mean square error and the R ² coefficient. When compared to ANN, the findings indicated that adaptive neuro-fuzzy inference system (ANFIS) displayed a much greater capacity to predict real-time flash floods [119]. A large number of ANNs were designed and the results were compared. In order to teach the researchers, they picked water level data from various stations between 1970 and 1985, and the data from 1986 to 1987 were utilized for verification. The ANNs performed well, even in cases when the dataset was not full, and they were a feasible option for making correct predictions. ANNs introduced the prospect of lowering analytical expenses by decreasing the amount of time spent on data analysis that was previously required [120].

Figure 13

General diagram of ANN [121].

4.3 Ensemble prediction systems

Ensemble learning basic idea is to reduce the probability of inference mistakes caused by a single classification model. Although the various classification models employed in ensemble learning have similar training outcomes and levels of accuracy, they also have distinct generalization capabilities, or inferential skills about diverse samples. The outcomes of these separate classification models are eventually integrated to provide the final classification performance, reducing the likelihood of misclassification dramatically. Furthermore, in ensemble learning, integrating the multiple inference outputs of many classifiers may not always lead to a good classification than employing the best particular classifier in the ensemble. Ensemble learning, on the other hand, has a high possibility of lowering the risk of selecting an especially terrible choice and improving overall classification prediction reliability [122]. Breiman introduced the bagging (bootstrap aggregating) ML ensemble technique in 1994. The bagging procedure divides the primary training dataset into numerous training data subsets (bootstrapped datasets), each of which is expected to build a classifier independently. To acquire the ultimate classification performance, the forecast results obtained by different classifiers are then combined via casting as shown in Figure 14. With a solid base in flood modeling, a variety of ML modeling alternatives are offered. As a result, there is a growing trend to move away from a single prediction model and toward an ensemble of models tailored to a certain purpose, cost, and database. ML ensembles are made up of a limited number of different models that provide greater flexibility than the alternatives. In flood prediction, ensemble ML algorithms have a lengthy history. Ensemble prediction methods have become popular in recent years (EPSs) [123]. The benefit of EPS is that it manages and evaluates ensemble techniques in a rapid and appropriate model. As a result, the functionality of EPS, particularly for flood modeling, may be enhanced [124]. EPSs may leverage numerous quick or statistical methods as classifier ensembles, such as ANNs, MLPs, DTs, rotation forest (RF) bootstrap, and boosting, enabling for greater accuracy and resilience in their classification of data. Based on the prediction rate employed in the event, the following ensembles prediction methods may be used to assess the likelihood of flooding. The methods for generating these EPS projections are quite simple. Some practical EPSs are built on a Monte Carlo structure of NWPs, with one realization based on a “central” assessment (the control prediction) and others depending on troubling the baseline circumstances (the perturbed forecasts). Based on the prediction site, the range of clusters ranges from 10 to 50 [125]. It is worth noting that tiny improvements in rainfall data are really not noticeable, which is terrible for hydrology [124].

Figure 14

Flow chart of Bagging (Bootstrap aggregation) [122].

4.4 SVM

Statistical process of learning has been used to develop and classify the SV as a nonlinear search method [126]. Subsequently, the SVM was developed as a kind of SV that was intended to lower the predicted error of learning algorithms by minimizing over-fitting. This method is an iterative machine learning that relies on the statistical learning theory and the structural risk reduction rule [111]. The ideal hyperplane performs a policy function in binary integer SVM for classification among two classes related to particular data. There are two approaches to achieving optimum original training categorization. To obtain a complete separation among training sample classes, hard boundary maximization might be used as shown in Figures 15 and 16. The hyperplane conditional probability is used to determine the peak value. It aids in increasing the length between both the hyperplane and the closest training relevant data. After determining the hyperplane, a different dataset may be fed into the SVM classifier. Depending on the direction of the input in relation to the test set, class is given +1 or 1 [2] as shown in Figure 16, in the event of a classification method with many classes. It is often used in flood modeling. SVM’s training process creates algorithms that generate additional non-probabilistic linear programming classifications that, via reverse conflict resolution, reduce the empirical classification error, and maximize the visual gap [127]. SVM is a ML algorithm that predicts a number in the future based on previous data [128]. The SVM has been expanded as a regression technique, termed as SVR, during the last two decades. SVM is still very much in inception in hydrological models, despite its popularity in many other domains. There have been many contradicting outcomes in its use so far. Scientific research shows that SVMs have similar over-fitting and under fit difficulties, although the excessively problem is more detrimental than the under-fitting issue, which has never been well tackled by the scientific community [129]. This is a highly good technique of learning about an SVM model’s function and flaws by seeing how it reacts to varying amounts of rainfall. For now, we do not know what this crucial response characteristic means for SVM’s evolution in hydrology, but future study will surely increase our understanding of this problem and allow models developed to make better use of this data [130].

Figure 15

Binary classification of SVM [2].

Figure 16

Multi-classification of SVM.

4.5 WNN

Use the wavelet transform (WT) to extract information from multiple data sources by looking at the local fluctuations in time-series data. WT, in reality, has a major impact on simulation results [131]. The dependable decomposition of an original time series by means of a transform helps to enhance the value of the evidence. DWT, which decays the actual information into zones, improves flood forecast lead times by increasing the accuracy of predictions. When developing a model, better information may be extracted using DWT’s decomposition of the basic dataset into distinct resolution levels [132]. DWTs are commonly employed in flooding estimate time because of their advantageous qualities. Rainfall–runoff DWTs were frequently used in flood simulation. Using two large-scale climatic data like SST and SLP as variables, a WNN hybrid model can estimate maximum monthly discharge of the Madarsoo watershed in northern Iran [133]. The achievement of the prediction of highest monthly release of three different hydrometry terminals of the drainage basin was measured using error metrics such as root-mean-square error [134], mean absolute error, and similarity measures such as coefficient of correlation (R), and Nash–Sutcliffe coefficient [135]. To forecast flood discharges in the period from March to August, the WNN hybrid ML model greatly exceeded the solo ANN model and MLR model in all situations [136]. To enhance environmental monitoring equipment, mechanized weather station networks capable of collecting minimum temperature and rainfall data have been installed almost globally since the turn of the century. Thus, there are now many temperature-dependent records available from many places, with few gaps and lasting over a decade [137]. As a result of recent research trying to combine ANN-based hydrology to wavelet analysis, the purpose of this study is to develop and validate various hybrid WNN models to forecast future monthly rainfall in Andalusia (Southern Spain) using only brief thermo-precipitation valid and reliable data sources [126,137]. The ability to predict river flow with high accuracy is critical in lowering the danger of flash flooding. It has become increasingly obvious that data-driven and hybrid strategies are being used to address the nonlinear and changeable aspects of hydraulic systems. As a result, this research introduces a unique hybrid WNN technique for river flow forecasting that includes image enhancement. In order to do this, the wavelet approach is used to evaluate the data that have been acquired as shown in Figure 17.

Figure 17

Wavelet model for reducing flood risk [138].

4.6 ANFIS

The fuzzy logic is a descriptive modeling system that employs a soft computing method that is based on natural language to get its results. Fuzzy logic is a reduced mathematical formula that relies on adding expert information into a fuzzy inference system (FIS) to make decisions in uncertain situations. Additionally, an FIS simulates human learning by using an approximation function that is less complicated than the original function, which has considerable promise for nonlinear modeling of severe hydrological processes, notably floods. For example, researchers looked at river level predictions utilizing a GIS, as well as rainfall–runoff models for water level. Neuro-fuzzy is a more sophisticated kind of fuzzy-rule-based modeling that combines the BPNN with the commonly used least-square error approach to produce a hybrid version. The Takagi–Sugeno (T–S) fuzzy modeling approach, which was developed via the use of neuro-fuzzy clustering, is also frequently used in the RFFA industry. ANFIS is a more sophisticated type of neuro-fuzzy that is based on the T–S FIS, which was initially defined in the 1970s. ANFIS is widely regarded as one of the most trustworthy estimators for complicated systems in use today. ANFIS technology, which combines ANNs and fuzzy logic, delivers more learning power. Specifically, this hybrid ML technique relates to a collection of sophisticated fuzzy rules that are ideal for modeling nonlinear flood dynamics. An ANFIS operates by using neural learning algorithms to detect and tune the variables and structure of an FIS. This is how it works. With the use of ANNs, the ANFIS hopes to identify any missing fuzzy sets in the database. ANFIS has become quite prominent in flood modeling because of its quick and simple application, precise learning, and excellent generalization capabilities, among other factors. While BPNNs are now commonly employed in this field, the MLP, a more sophisticated approximation of ANNs, has lately grown in favor as a replacement for them.

4.7 Multilayer perceptron

It is a kind of FFNN that uses supervised learning of BP for training a network of linked nodes of various layers, and it is one of the most often used. The MLP is characterized by its flexibility, nonlinear operation and a significant number of layer combinations. The approach was extensively utilized in flood forecasting and other sophisticated hydrogeological simulations as a result of these properties. MLP models were shown to be more efficient than ANN models in a study of ANN classes being used for flood simulation, as well as having more generalization ability. In spite of this, the MLP is often regarded as more difficult to optimize. Back-percolation training techniques are used to determine the transmission mistake in hidden network nodes on an individual level, allowing for a more sophisticated modeling framework. It is interesting to note that the MLP has acquired levels that are high within hydrologists than any other variety of ANNs (e.g., FFNN, BPNN, and FNN) in recent years. Furthermore, because of the large number of case studies that used the conventional form of MLP, it deviated from the usual form of ANNs. Furthermore, writers of works in the field of flood prediction that makes use of the MLP designate to their methods as MLP models when referring to their models. This was taken into consideration when we determined to create a distinct area for the MLP.