Leveraging machine learning and open-source spatial datasets to enhance flood susceptibility mapping in transboundary river basin

ABSTRACT

Floods pose devastating effects on the resiliency of human and natural systems. flood risk management challenges are typically complicated in the transboundary river basin due to conflicting objectives between multiple countries, lack of systematic approaches to data monitoring and sharing, and limited collaboration in developing a unified system for hazard prediction and communication. An open-source, low-cost modeling framework that integrates open-source data and models can help improve our understanding of flood susceptibility and inform the design of equitable risk management strategies. This study integrates open-source datasets and machine -learning techniques to quantify flood susceptibility across the data-scare transboundary basin. The analysis focuses on the transboundary Gandak River Basin, spanning China, Nepal, and India, where damaging and recurring floods pose serious concern. flood susceptibility is assessed using four widely used machine learning techniques: Long-Short-Term-Memory, Random Forest, Artificial Neural Network, and Support Vector Machine. Our results exhibit the improved performance of Artificial Neural Network and Support Vector Machine in predicting flood susceptibility maps, revealing higher vulnerability in the southern plains. This study demonstrates that remote sensing and machine learning can improve flood prediction, hazard mapping, and susceptibility analysis in a data-scare environment.

KEYWORDS:

• Transboundary river basin
• susceptibility analysis
• open-source dataset
• machine learning
• flood risk management

1. Introduction

Floods, characterized by river waters overflowing into nearby floodplain areas, present widespread natural calamities that inflict significant human and economic losses (Pham et al. 2021). A surge in extreme climatic events, such as intense rainfall, has led to a substantial increase in flood occurrences, primarily attributed to human-induced climatic factors (Yin et al. 2018). Activities such as altering river basins, changing land use patterns, and climate variability have extensively impacted runoff generation, elevating flood risk (Merz et al. 2021). While precipitation typically drives runoff in a catchment, it undergoes modifications influenced by catchment characteristics, such as topography, land use changes, and river network modifications. The cumulative impacts of these factors trigger larger and more severe floods (Merz et al. 2021). The global economic toll of floods between 1980 and 2013 surpassed $1 trillion, with over 220,000 lives lost (Winsemius et al. 2016). In the twentieth century alone, global flood-related fatalities exceeded seven million, with an annual average loss of over $104 billion (Merz et al. 2021). These losses extend to direct asset damage from natural disasters, totaling over $300 billion (Rentschler, Salhab, and Jafino 2022). Flood risk is particularly pronounced in countries with low economic growth, such as Nepal, where there is a dire need for comprehensive studies, preventive measures, and mitigation strategies (Rentschler, Salhab, and Jafino 2022). Furthermore, Earth system models predict a notable increase in flooding frequency in significant parts of South and Southeast Asia (Hirabayashi et al. 2013). The impact of this projection is exemplified by the estimation that one-third of total flood-related displacements in Asia and the Pacific region occur in South Asia, accounting for 61.4 million people between 2010 and 2021, with 44 million affected solely in Nepal and India (Anzellini et al. 2022).

Floods transcend political boundaries and affect many nations. Approximately 40% of the world's population resides in transboundary watersheds, emphasizing the necessity for cross-border policies that integrate water and disaster management (Cooley and Gleick 2011; Pandey et al. 2023). Transboundary water resources, particularly in the face of global change, including climate variability, pose complex challenges (Munia et al. 2020). Managing these resources is complicated by the mismatch between political and watershed boundaries despite numerous treaties and agreements aimed at dispute resolution (Cooley and Gleick 2011). Floods often emerge as a major source of friction in transboundary river basin management, as water respects no borders (Bakker 2009). In addition, several restrictions on data sharing further complicate the development of flood models, communication of hazards, and the formulation of risk management strategies.

In the South Asian context, where China, Nepal, and India share strong cultural, economic, and human resource ties, the Himalayan-Ganges Basin is a major disaster hotspot (Gupta et al. 2021). These countries face the common challenge of flood disasters, particularly along shared borders (Mehravar et al. 2023). Around 6000 rivers from Nepal are flowing into India, covering an extensive 45,000 km. Flood risks affect 80% of the total area and 97% of the population, making them a critical concern for both nations (Sharma, Fu, and Kattel 2023). The major drawbacks of flood studies are the lack of data and knowledge sharing in major transboundary rivers. Effective flood control measures demand cooperation among China, Nepal, and India to mitigate the impact of floods on lives and property. Most damage occurs in the plains near the Nepal-India border, particularly in Terai in Nepal and Bihar/Uttar Pradesh in India, which are regions highly vulnerable to severe floods, except for some flash floods and Glacial Lake Outburst Floods (GLOFs) near the China–Nepal border (Khanal et al. 2015). Despite their water-rich environment, these regions host some of the poorest communities at risk (Gupta et al. 2021). Achieving flood resilience in these areas necessitates coordinated efforts between Nepal and India, considering their extensive river boundary; that is, 595 km out of 1,808 km of the border are rivers (Adhikari 2013). Among 191 high or very high-risk lakes with potential GLOFs, most lie in the Sino-Nepal Border, where data and information sharing are crucial. This transboundary region is a vital economic area between Nepal and India; however, various constraints affect timely data sharing, communication, and coordination, adversely affecting disaster preparedness (Zheng et al. 2021).

In Nepal and India, existing treaties and committees have primarily focused on water and power sharing, overlooking critical aspects such as the effects of climate change on transboundary populations (Dixit and Shukla 2022). Timely amendments of treaties or agreements based on these advanced scientific approaches can benefit the state to local-level decision-makers. They can make decisions in a win-win situation (Cooley and Gleick 2011). Nepal and India have numerous agreements and treaties related to transboundary rivers, of which Koshi, Gandak, and Mahakali are the major ones. In the case of Nepal-India agreements, the focus is mainly on constructing engineering structures and sharing water and hydroelectricity. There are limited efforts to assess the critical challenges, such as climate change and related effects on human and natural systems in transboundary basins. In the Gandak River Basin (GRB), frequent flooding has pronounced effects on major downstream areas of the Gandak Barrage. Majorly, the parts of Nawalpur, Parasi, Kapilvastu, and Rupandehi districts of Nepal and the northeastern part of Uttar Pradesh of India are flooded regularly. Heavy rainfall upstream of the southern plain of the basin and the need for maintenance of embankments, siphons, and canals by the authorities have also caused flooding downstream of the barrage (Dixit and Shukla 2022).

Studies show that losses are relatively high at the transboundary basins, where the institutional capacities are constrained (Bakker 2009; Cooley and Gleick 2011). Detailed scientific studies are hindered by constraints related to data availability and geopolitical influences; however, proper management decisions based on scientific studies can only provide a realistic approach. Flood susceptibility mapping is pivotal in floodplain management, aiding informed decision-making regarding flood hazards (Duwal, Liu, and Pradhan 2023; Ghosh, Saha, and Bera 2022). Floods can trigger secondary hazards such as erosions and landslides, necessitating the quantification of susceptible areas using advanced modeling techniques (Mehravar et al. 2023; Panahi et al. 2022). Various categories of flood susceptibility mapping exist, including physical modeling, statistical modeling, multi-criteria decision analysis, and machine learning techniques (Mehravar et al. 2023; Pham et al. 2021). Although physical modeling excels in smaller areas, data acquisition and computational complexities for large areas present challenges for such models (Riazi et al. 2023). On the other hand, statistical modeling, though versatile, struggles to represent the nonlinear nature of floods due to its linear approach. This limitation often results in underestimating flood extent due to oversimplification (Zhou et al. 2022).

Machine learning, a data-driven approach, offers a promising alternative for flood susceptibility mapping. It utilizes flood inventory data and conditioning factors to create models capable of identifying risk areas. Field surveys, traditionally essential for understanding floods, are expensive and time-consuming (Bui et al. 2015; Riazi et al. 2023) and can be entangled into administrative and transboundary rules and formalities. However, advancements in remotely sensed data acquisition and geospatial data accessibility have paved the way for machine-learning applications in flood analysis (Chapi et al. 2017; Chen et al. 2019; Khosravi et al. 2020; Costache et al. 2020; Schmidt et al. 2020; Towfiqul Islam et al. 2021; Avand et al. 2022a). The ability of machine learning to handle fluctuating data derived from climate models and human activities is a vital tool for future flood prediction (Schmidt et al. 2020). Unlike physics-based models, which can be computationally expensive for real-time forecasting, machine-learning models can offer a rapid and robust approach to flood susceptibility detection (Zhou et al. 2022). Over the past decade, the use of machine learning in hazard characterization across a wide range of spatial and temporal scales has drastically increased (Arora et al. 2021; Pham et al. 2021; Avand et al. 2022b; Antzoulatos et al. 2022; Saber et al. 2023; Devitt et al. 2023).

This study endeavors to bridge the knowledge gap by applying machine learning approaches to transboundary flood management, shifting from the prevailing state-centric framework using open-source datasets and low-cost modeling approaches. Leveraging open-source data and machine learning methodologies circumvents the significant challenges posed by limited data availability and methodological intricacies in disaster-related research within developing countries such as Nepal. This study was thus focused in developing a robust data-driven predictive model that enhances flood prediction accuracy, serving as a foundation for proactive flood hazard management across the entire river basin. The study area includes the GRB, a shared concern between China, Nepal, and India. Past approaches have primarily focused on structural measures; however, the dynamic nature of floods in transboundary communities calls for a more scientific approach. By leveraging machine learning, this study aims to contribute valuable insights that can inform strategies for preparedness, resource allocation during flood events, and effective mitigation planning, ultimately minimizing the impact of future floods on river basin communities.

2. Study area

GRB is in the transboundary region between China, Nepal, and India (Figure 1). The GRB extends between the Koshi and Karnali Rivers, spanning an area of 34,960 km² across these three countries. Seven tributaries feed the Gandak or Narayani River: Trishuli, Budhi Gandaki, Marsyangdi, Seti, Daraundi, Madi, and Kali Gandaki (Dixit 2017). The principal river within the GRB, the Kali Gandaki, originates in the Mamang Bhot Tibetan plateau, located north of the Annapurna and Dhaulagiri Mountain ranges (Dixit 2017; JICA 1999). It carves through these mountain ranges, forming the world's deepest gorge. Upstream of Devghat in Nepal, six major and numerous smaller waterways from Nepal's Middle Mountain region converge with the Kali Gandaki, collectively referred to as Sapta Gandaki or Narayani in Nepal. After flowing through the Chitwan Valley, the river turns westward before heading south through the Siwalik range and crossing into India at Bhaisalotan. In Chitwan Valley, the East Rapti River and Riu Khola join the Narayani before it enters India. The river traverses the Chure-daunne hills, West Champaran of Bihar, and ultimately influences the Ganga River (Dixit 2017; Panthi et al. 2015). The GRB exhibits significant variations in physiography, climate, and vegetation from north to south, encompassing regions such as the Tibetan Tethys Himalaya, higher Himalaya, lesser Himalaya, Siwaliks, and Terai (Panthi et al. 2015).

Figure 1. a) The prepared inventory of flood and non-flood points in the Gandak River Basin, b) The administrative map representing the extension of Gandak River Basin from China, Nepal to India, c) The Ganges basin representing Gandak River Basin as one of the tributaries.

Glaciated valleys with extremely rugged and steep terrain occur in the Himalayas, where the average elevation is 3000 m above sea level. The abundance of hills with larger areas of mixed forests and terraced farming is an identifying characteristic of the lesser Himalayas. Meanwhile, the composition of highly erodible soil and landslide occurrence indicate the Siwalik regions of the GRB. Terai region is low-lying flat land that is flooded yearly. The climatic behavior of the GRB is dominated mainly by monsoonal rain occurring from mid-June to late September each year (JICA 1999; Panthi et al. 2015). Snow and precipitation control the glacier-fed river system (Nandi, Srivastava, and Shah 2017). The region had a maximum daily discharge of 14,100 m³/s, which was recorded on August 5, 1974, on the Kaligandaki River. The upstream portion of GRB has a subalpine or temperate climate. In contrast, subtropical to temperate and more tropical climates are observed in the mid and downstream part of GRB, respectively. Previous studies showed an increasing trend of climate indices on warm days and a decreasing trend on colder days (Bajracharya, Acharya, and Ale 1970; Shrestha et al. 1999).

3. Methodology

The methodology involved in the study is shown in Figure 3. The applied procedure in the determination of flood susceptibility involves: a) Preparation of the flood inventory, b) Preparation of the flood conditioning factors, c) Selection of the suitable conditioning factors through multicollinearity test, d) Development of flood susceptibility maps, e) Assessment of the accuracy of the models f) Generation of the flood susceptibility map using a machine learning algorithm.

Figure 2. Flood conditioning factors used in the study.

3.1. Flood inventory map

In general, flood susceptibility mapping is the binary classification method that involves the use of 1 for the flood location and 0 for the non-flood location (Choubin et al. 2019; Tehrany, Pradhan, and Jebur 2014; Tehrany, Pradhan, and Jebur 2015; Tehrany et al. 2015; Towfiqul Islam et al. 2021). The flood point represents the precise location where flood has occurred in recurrent years, and the non-flood point represents the location where flood has not been recorded for the last ten years, including the areas of hills and mountains where the impact of the flood is low(Khosravi et al. 2016). Developing a flood inventory map is the initial stage for preparing a flood susceptibility map (Choubin et al. 2019). Accurate assessment of flood susceptibility requires precise flood locations for estimation of potential occurrences in the future, as flood susceptibility mapping is predictive analysis (Choubin et al. 2019; Tehrany, Pradhan, and Jebur 2015; Tehrany et al. 2015; Towfiqul Islam et al. 2021). Flood inventory mapping provides insights into the characteristics of historical flood events and information regarding inundated areas (Wang et al. 2020). This study developed flood inventory based on the methodological chart presented (Figure 2).

In the initial phase of flood inventory compilation, national flood archives were acquired through the Disaster Risk Reduction (DRR) portal. Additionally, an exhaustive search was conducted using English and Nepali keywords to scrutinize online news sources for flood-related information, selected sources are listed in supplementary Tables S1 and S2. The documentation process involved recording the geographical locations and temporal occurrences of floods throughout the monsoon period, spanning from July to September. In QGIS platform, the point features were created in the reported location of the flood. ‘Flood’ field was created and flooded location was given 1 (positive) value. Concurrently, global flood datasets were obtained from the flood observatory, sourced from the flood hazard map accessible at the Department of Hydrology (DFO) website (https://floodobservatory.colorado.edu). The point features were created in the represented location of the flood as well as non-flooded locations. The ‘Flood’ field was created, and the flooded location was given a 1 (positive) value, and the non-flooded location was given a 0 (negative) value.

Normalized Difference Water Index (NDWI) is the spectral index that separates water bodies from land using spectral characteristics (Nandi, Srivastava, and Shah 2017). However, some previous studies (Munasinghe et al. 2018; Nandi, Srivastava, and Shah 2017) have suggested using modifiedNDWI as the built-up area noises the water surface extracted from NDWI. Furthermore, the high bulk of floodwater sediments causes trouble representing it with normal indices. flooded areas were generated using Google Earth Engine (GEE). The sentinel-1 Level-1 GRD (Ground Range Detection) images for different dates of floods were filtered. The initiation of the image filtration process involved the meticulous selection of polarization ‘VV,’ coupled with subsequent considerations for pass direction, spatial resolution, and confinement to the designated study area. The preprocessing of Sentinel-1 images, accessed through Google Earth Engine (GEE) libraries, encompassed essential procedures such as the application of orbit files, removal of border noise through Analysis Ready Data (ARD), thermal noise mitigation, radiometric correction, terrain correction, and conversion of the backscatter coefficient to decibels. a smoothing filter was applied to the generated image to address the intrinsic speckle effect inherent in radar imagery. The determination of an optimal threshold, set at 1.24, was achieved through iterative experimentation, leading to the segmentation of the raster derived from the change-detection algorithm into distinct categories denoting flooded (1) and non-flooded (0) segments.

Areas characterized by seasonal water coverage were discerned through an analysis of the seasonality patterns within the Global Surface Water Dataset. Concurrently, areas subject to water coverage for a duration exceeding ten months were identified and masked accordingly. The Hydroshed SRTM dataset was employed to eliminate regions characterized by slopes surpassing 5%, and noise reduction was executed based on the connectivity of flood pixels, specifically those linked to eight or fewer neighboring pixels. Subsequently, the derived raster representing the extent of flooding was transformed into a polygon format and exported for subsequent stages of analysis and processing. Again, ‘Flood’ field was created and flooded location was given 1 (positive) value and non-flooded location was given 0 (negative) value. The obtained points from the mentioned source from Table 1 are scrutinized for repetition and final flooded and non-flooded points were derived as represented in Figure 1(a).

Table 1. Sources utilized in the preparation of flood inventory.

Data	Spatial Resolution	Source
MNDWI	10 m x 10m	(Sentinel 2013)
Flood Extent	10 m x 10m	(Tazmul Islam and Meng 2022)
Seasonal Water Coverage	30 m x 30m	(Pekel et al. 2016)
Flood dataset	30 m x 30m	(Brakenridge 2016)

3.2. Flood conditioning factors

The conditioning factor for flood susceptibility mapping may vary based on the watershed characteristics. Due to the flood's dynamic nature, including different factors for flood conditioning is necessary (Ali et al. 2020). Based on the previous studies performed by various researchers, factors like slope, aspect, curvature, distance to a river (DTR), drainage density (DD), elevation, rainfall, sediment transport index (STI), stream power index (SPI), topographical wetness index (TWI), soil type, normalized difference vegetative index (NDVI) and land use/land cover (LULC) are utilized in this study (Arora et al. 2021; Chapi et al. 2017; Chen et al. 2019; Choubin et al. 2019; Khosravi et al. 2019; Shafapour Tehrany et al. 2019; Shafizadeh-Moghadam et al. 2018; Tehrany, Pradhan, and Jebur 2014; Tehrany, Pradhan, and Jebur 2015; Tehrany et al. 2015) represented in Figure 3. SRTM DEM was downloaded from the USGS database and obtained a Landcover map from ESRI using GEE. rainfall data was extracted from the IMERG product of Global Precipitation Measurement and Soil Data was derived from Harmonized Soil Database from the Food and Agriculture Organization of the United Nations.

Figure 3. Schematic diagram representing data-model integration framework for flood susceptibility analysis.

The slope measures the angle between the terrain and the horizontal datum (Tehrany et al. 2015). A steep slope in the watershed enhances the speed of the surface runoff, as it reduces the infiltration time of the rainfall (Gudiyangada Nachappa et al. 2020; Shafapour Tehrany et al. 2019). Aspect is the direction of the ultimate slope of the terrain surface (Choubin et al. 2019). Curvature represents the shape of the terrain and affects runoff and infiltration of the area, causing floods (Chapi et al. 2017; Khosravi et al. 2019; Mehravar et al. 2023). Generally, the curvature is divided into three classes: concave, convex, and flat. The proneness of runoff is seen in the convex surface, which is eventually responsible for down-slope flooding (Khosravi et al. 2019). DTR plays a significant role in flooding as it influences the velocity and extent of the flood (Tehrany et al. 2015). Nearby areas to the river are greatly affected by the overflowing of river banks (Chapi et al. 2017), signifying the closeness or proximity of the main river channel (Shafizadeh-Moghadam et al. 2018). DD is the stream's total length divided by the catchment's total area (Horton 1945; Shafizadeh-Moghadam et al. 2018). Previous studies (Dekongmen et al. 2021; Horton 1945; Ignacio and Walling 1968; Mashao et al. 2023; Pallard, Castellarin, and Montanari 2009) have shown that the catchment with higher stream density demonstrates a rapid response to rainstorms and high flooding potential. The elevation is considered the influential controlling factor in the case of flood susceptibility mapping. In general, areas of higher elevation are less prone to flood than flat lands as water flows from the confined channels in higher mountains to lower terrain and spreads, causing a higher risk of flood (Khosravi et al. 2019; Riazi et al. 2023; Tehrany, Pradhan, and Jebur 2015; Tehrany et al. 2015).

Rainfall is the primary source of runoff generation for flooding (Riazi et al. 2023). An increment in rainfall enhances the probability of flood occurrence (Chapi et al. 2017; Shafizadeh-Moghadam et al. 2018). Previous studies by Cao et al. (2016) have demonstrated the potential of higher flooding in the lower plain due to heavy rainfall in the upper area. The STI is also a major flood conditioning factor (Chapi et al. 2017; Fang et al. 2021; Shafapour Tehrany et al. 2019). It signifies the impact of topography on erosion and indicates the strength of sediment displacement caused by the flow of water (Fang et al. 2021; Shafapour Tehrany et al. 2019). The SPI represents the erosion power of flowing water (Gudiyangada Nachappa et al. 2020; Riazi et al. 2023; Tehrany, Pradhan, and Jebur 2015; Tehrany et al. 2015). The TWI represents the potential occurrence of wetness or degree of saturation of soil in the study area (Fang et al. 2021). The soil layer acts as the critical factor in triggering the flood as it controls infiltration and erosion status in the study area (Riazi et al. 2023). The permeability of various lithological units affects their behavior towards runoff generation. Geology also plays a crucial role in the formation of drainage patterns, which in turn influence the development of floodplains (Riazi et al. 2023). NDVI is the representation of vegetation density. Past studies have revealed the negative relationship between NDVI and flooding of the area (Mehravar et al. 2023). LULC is also an important influencing factor, as the densely populated vegetative regions are less prone to flooding (Tehrany, Pradhan, and Jebur 2014). Thick vegetation decelerates the water flow, whereas bare land facilitates runoff generation Table 2.

Table 2. Data utilized in the development of flood conditioning factors.

Conditioning Factors	Spatial Resolution	Primary source	References
Elevation	30mx30m	DEM	(NASA JPL, 2013)
Slope	30mx30m	DEM	(NASA JPL, 2013)
Aspect	30mx30m	DEM	(NASA JPL, 2013)
Curvature	30mx30m	DEM	(NASA JPL, 2013)
Distance to River	30mx30m	DEM	(NASA JPL, 2013)
Drainage Density	30mx30m	DEM	(NASA JPL, 2013)
Sediment Transport Index	30mx30m	DEM	(NASA JPL, 2013)
Stream Power Index	30mx30m	DEM	(NASA JPL, 2013)
Topographic Wetness Index	30mx30m	DEM	(NASA JPL, 2013)
Soil Type	30’ x 30'	HWSD	(FAO 2012)
Normalized Difference Vegetation Index	30mx30m		(Sentinel 2013)
Land Use/Land Cover	10mx10m		(Karra et al. 2021)
Rainfall	0.1° x 0.1°		(Huffman 2015)

3.3. Machine learning algorithm

3.3.1. Random Forest (RF)

RF model involves the use of two procedures for prediction generation. 1): ‘Bagging’ introduced by Breiman (Richman and Wüthrich 2020), which comprises the generation of multiple versions of a predictor using the training set, and the numerical outcome of the prediction is done through averaging over different versions. Likewise, 2) Another part of the RF model includes ‘random selection of the governing features.’ These two features enable the RF model to generate the samples and replace and generate multiple regression trees for training. The output is forecasted based on the voting involving multiple classifiers (Towfiqul Islam et al. 2021).

3.3.2. Support Vector Machine (SVM)

Support Vector Machine (SVM) is the supervised learning classifier that utilizes a binary classification method following the structural risk minimization principle (Tehrany, Pradhan, and Jebur 2014; Tehrany et al. 2015). SVM forms the hyperplane, which aids in the transformation of the nonlinear parameter into a simple and understandable linear parameter (Tehrany et al. 2015). A kernel function is a mathematical function applied to transform data in flood susceptibility generation (Towfiqul Islam et al. 2021). The training data is used to construct a hyperplane, which signifies the mapping of original input into the high-dimensional feature space (Tehrany et al. 2015; Towfiqul Islam et al. 2021). SVM involves determining the maximum margin of separation, and the hyperplane is built on the center of the maximum separation. If the position of the point is on the hyperplane, it is labeled as +1, or else it is labeled as −1. The training points found to be near best-fit hyperplanes are called support vectors. With the completion of the construction of the hyperplane, the new data can be classified based on its position relative to the hyperplane (Tehrany, Pradhan, and Jebur 2014).

3.3.3. Artificial Neural Networks (ANN)

ANN involves using connections made through neurons, namely hidden and output neurons. ANN utilizes a multilayer perceptron having multiple hidden layers where the multilayer perceptron consists of a weighted sum of the input variables (Dtissibe et al. 2020). ANN serves better than many other methods involving statistics in hydrological analysis, but database length could cause hindrances in the modeling process. Likewise, errors in ANN can be observed while forecasting the value outside the training dataset provided (Tehrany, Pradhan, and Jebur 2014).

3.3.4. Long Short-Term Memory (LSTM)

LSTM (Hochreiter and Schmidhuber 1997) is a Recurrent Neural Network (RNN) type that captures long-term dependence and sequential data patterns. The key characteristic of LSTM is its ability to maintain a memory state that can retain information for long periods, allowing the network to remember and forget information selectively over time (Chen et al. 2023; Fang et al. 2021; Liu et al. 2020; Sherstinsky 2020). The major shortcoming of conventional RNN is the problem of gradient diminishing or exploding, which LSTM controls. It is an improvised RNN that discards useless information and remembers key information. It is powerful in learning the complex relationship between data, so it is a good tool for improving the accuracy of flood susceptibility mapping (Fang et al. 2021).

3.4. Model evaluation criteria

Statistical measures are effective tools for evaluating the performances of the developed models. Evaluation of the training dataset reveals the fitness of the dataset to be used, while assessment of the testing dataset reveals the model's predictive capacity (Khosravi et al. 2019). Statistical measures like Specificity, Sensitivity, Accuracy, Kappa Score, Mean Squared Error (MSE), and Root Mean Squared Error (RMSE) are determined based on true positive, false positive, true negative, and false negative values (Luu et al. 2021). The confusion matrix (CM) of size is generated in the sequence represented in the equation given below. The number of pixels classified correctly as flood pixels are true positive (TP), and those classified correctly as non-flood pixels are true negative (TN). Similarly, false positive (FP) represents the number of pixels classified incorrectly as flood, and false negative (FN) represents the number of pixels classified incorrectly as non-flood (Chapi et al. 2017). $\mathrm{CM}=\left[\begin{array}{cc}\mathrm{TP}& \mathrm{FP}\\ \mathrm{FN}& \mathrm{TN}\end{array}\right]$ $\mathrm{Specificity}={\displaystyle \frac{\mathrm{TN}}{\left(\mathrm{TN}+\mathrm{FP}\right)}}$ $\mathrm{Sensitivity}={\displaystyle \frac{\mathrm{TP}}{\left(\mathrm{TP}+\mathrm{FN}\right)}}$ $\mathrm{Precision}={\displaystyle \frac{\mathrm{TP}}{\left(\mathrm{TP}+\mathrm{FP}\right)}}$

3.5. Receiver Operating Characteristics (ROC)

ROC curves are utilized to evaluate the performances of FSM models developed using artificial intelligence and validate the results obtained with the validation dataset (Gudiyangada Nachappa et al. 2020; Wang et al. 2020). The ROC curve is constructed by plotting sensitivity on the Y-axis and 1-specificity on the X-axis. In general, the area Under the Curve of Receiver Operating Characteristics (AUROC) value obtained represents the spatial relation between the trained data and testing data. It is the quantitative index representing the prediction ability of the model to categorize flood occurrence or non-occurrence (Tehrany, Pradhan, and Jebur 2015; Chapi et al. 2017; Bui et al. 2018b; Wang et al. 2020). AUROC value ranges from 0.5–1, where 0.5 represents the inaccurate model, and 1 illustrates the perfect model (Bui et al. 2018a; Wang et al. 2020).

3.6. Feature selection criteria

Selecting suitable features is necessary for developing an efficient machine-learning model. The involvement of features having predictive power results in proper prediction, whereas some features may generate noise and reduce the overall performance of the developed models (Chapi et al. 2017; Dodangeh et al. 2020). Statistical approaches are selected to perform the selection of suitable features. For the quantification of influential factors, the Information Gain Ratio (InGR) is adopted (Towfiqul Islam et al. 2021). The greater the value of InGR, the higher the factor's influence on the target variable (Dodangeh et al. 2020). Similarly, several published articles suggest the application of the Variable Inflation Factor (VIF) and Tolerance (TOL) for the evaluation of collinearity dependence (Arora et al. 2021). It is recommended that a model to be free from the multicollinearity tolerance should have a value greater than 0.1, and VIF should have a value less than 10 (Arora et al. 2021; Mehravar et al. 2023).

4. Results

4.1. Adoption of conditioning factors

Multicollinearity and information gain ratio (IGR) tests are performed to test the suitability of the conditioning factors. Several published articles suggest the application of the Variable Inflation Factor (VIF) and Tolerance (TOL) for the evaluation of collinearity dependence (Arora et al. 2021). In this study, the variables were scrutinized for VIF and tolerance to detect multicollinearity in the multiple regression used for flood susceptibility generation. Furthermore, Pearson's correlation test was conducted to gain insights into the correlation between the flood and flood conditioning factors. It is recommended that a model to be free from the multicollinearity tolerance should have a value greater than 0.1, and VIF should have a value less than 10 (Arora et al. 2021; Mehravar et al. 2023). The initial assessment identified high VIF values for aspect, indicating collinearity. Subsequently, the aspect was removed, and the remaining ones were used for model training. Upon reevaluation, elevation exhibited the highest VIF (5.415), followed by slope (2.270), while soil had the lowest VIF (1.069) (Figure 4). Similarly, regarding tolerance, soil showed the highest value (0.935), and elevation had the lowest (0.184). the highest correlation values were observed for slope (0.751), distance to river (0.698), and elevation (0.687), indicating a strong correlation between these factors and flood occurrence. The IGR values are higher for slope(0.44), elevation(0.37) and DTR(0.29) and low values are observed for soil(0.01), STI (0.04), SPI(0.04) and curvature(0.05) (Figure 4). The values of IGR clearly show that the topographic factors, majorly slope, elevation, and distance to the river, are major drivers of flooding in GRB.

Figure 4. Selection of parameters using multicollinearity test, Pearson's correction and Information gain ratio.

4.2. Model training and validation

LSTM, RF, SVM, and ANN models were developed and validated using training and validation datasets. Previous studies have suggested the application of AUROC for the proper validation of the models developed (Pourghasemi, Pradhan, and Gokceoglu 2012; Tehrany, Pradhan, and Jebur 2014; Tehrany et al. 2015; Tehrany, Pradhan, and Jebur 2015; Chapi et al. 2017; Bui et al. 2018a, 2018b; Choubin et al. 2019; Dtissibe et al. 2020; Pham et al. 2021; Chen et al. 2023). The higher the value of the AUROC obtained, the higher the model's performance in the flood susceptible area detection (Towfiqul Islam et al. 2021). The ROC curve is the graph developed by plotting sensitivity on the X-axis and specificity on the Y-axis. Sensitivity or true positive rate is the measure for demonstrating how well the model can predict positive instances. Likewise, specificity or true negative rate indicates how well one can predict negative instances. Results demonstrated that all the models performed with higher precision (>0.90) AUROC was employed for model validation, with SVM demonstrating the highest AUROC (0.996) for training datasets, followed by RF (0.994), ANN (0.993), and LSTM (0.992) (Figure 5a). RF exhibited the highest AUROC (0.994) for testing datasets, followed by ANN (0.993), SVM (0.991), and LSTM (0.989) (Figure 5b). Other metrics, including MSE, RMSE, and Kappa score, were employed for comprehensive evaluation. Although the prediction evaluation showed the performance of models to be similar, the spatial pattern of the prediction maps is different, as presented in Figure 6. So, to reduce the uncertainty in spatial prediction, other metrics like root mean squared error (MSE), root mean squared error (RMSE), and Kappa score (Bui et al. 2018b) are employed, as represented in Table 3. Regarding MSE and RMSE, RF outperformed other models with the lowest values of 0.025 and 0.157, respectively, indicating superior accuracy in predicting flood susceptibility. The F1-score, a measure of the balance between precision and recall, was consistently high across all models, ranging from 0.97–0.98. This signifies the models’ robustness in correctly classifying positive and negative instances. The Kappa statistic, assessing agreement between predicted and observed classifications, demonstrated that RF achieved the highest agreement with a value of 0.951. Overall, the comprehensive evaluation of these metrics underscores the efficacy of RF in flood susceptibility mapping, supported by its remarkable accuracy, balanced precision-recall performance, and strong agreement between predicted and observed classifications.

Figure 5. Comparison and validation of the Machine Learning Approaches for FSM.

Figure 6. Representation of different areas under the variable scale of flood susceptibility under different machine-learning models (Abbreviations: VL = very low, L = low, M = medium, H = high and VH = very high).

Table 3. Performance parameters for different machine learning models.

Model	MSE	RMSE	F1-score	Kappa
RF	0.025	0.157	0.98	0.951
SVM	0.033	0.181	0.97	0.934
ANN	0.031	0.175	0.97	0.938
LSTM	0.045	0.213	0.95	0.909

4.3. Flood susceptibility map

After training and validating the models, flood susceptibility maps were generated for each. The ANN model emerged as the highest-performing model, producing detailed and accurate flood susceptibility maps based on geospatial factors. The prepared flood conditioning factors generated a flood susceptibility map from each model. Creating and organizing probability maps into different categories of flood susceptibility maps is essential for visualizing the spatial predictions of floods (Chapi et al. 2017; Tehrany, Pradhan, and Jebur 2014). The natural break method was adopted based on the previous studies (Chen et al. 2019; Choubin et al. 2019; Costache and Tien Bui 2020; Dodangeh et al. 2020; Luu et al. 2021) for the reclassification of the obtained flood susceptibility map into different classes, namely, very low, low, moderate, high, and very high represented in Figure 6. The natural break method was suitable for this research as it assists in grouping similar values and enhances the differences between the classes (Luu et al. 2021). The area under each class of the flood susceptibility map was obtained, as represented in Figure 6 and Figure 7. The study revealed, for the LSTM model, 20% of the area lies under the very highly susceptible zone, followed by 3.8% of highly susceptible, 2.8% of medium susceptible, 3.7% of low susceptible, and 69.7% of the very low susceptible area.

Figure 7. Flood susceptibility maps of selected areas (a1, a2, a3) predicted under the Long-Short-Term-Memory model (A,B,C), Artificial Neural Networks (D,E,F), Support Vector Machines (G,H,I), and Random Forest (J,K,L).

Similarly, for the ANN model, 13.1% of the area lies under the very highly susceptible zone, followed by 8.7% of highly susceptible, 9.0% of medium susceptible, 12.1% of low susceptible, and 57.1% of the very low susceptible area. Likewise, for the SVM model, 10.2% of the area lies under the very highly susceptible zone, followed by 7.6% of highly susceptible, 10.7% of medium susceptible, 17.2% of low susceptible, and 54.3% of the very low susceptible area. Lastly, for the RF model, 7.9% of the area lies under the very highly susceptible zone, followed by 4.6% highly susceptible, 5.6% medium susceptible, 13.4% low susceptible, and 68.5% very low susceptible area. Comparison between the models revealed the highest value of high (7861.96 km²) and low (27390.09 km²) susceptible area was detected by the LSTM model. RF, SVM, and ANN models demonstrated a similar pattern for the distribution of the flood in the region, while the LSTM model showed a slightly different pattern.

According to the prediction made by all the models, areas near the river system were more prone to flooding than others. Likewise, it can be seen from Figure 7 that the flood has impacted regions lying in the lower portion of the GRB to the maximum extent.

5. Discussion

Flood susceptibility mapping is one of the major steps in identifying flood-prone areas. Using publicly available open-source data and methodologies such as machine learning has been a game changer in flood studies. Data restrictions and scarcity are the major hindrances to flood-related studies in transboundary river basins. The current research addresses this issue and prepares a robust model for flood studies. The applicability of the conditioning factors and performance of the machine learning algorithms to determine flood susceptibility differs according to the chosen factors and complexity of the study area (Seydi et al. 2023; Tehrany, Jones, and Shabani 2019). This study thus assessed the applicability of the conditioning factors and their effects on the performance of the model using multicollinearity and information gain ratio. The accuracy assessment and validation of the model were performed using different evaluation parameters. Besides this, the validation of the model was checked based on historically observed flood areas from various sources and field observations.

5.1. Factors selection and their effects on flood modeling

Selecting suitable features is necessary for developing an efficient machine-learning model. The involvement of features having predictive power results in proper prediction, whereas some features may generate noise and reduce the overall performance of the developed models (Chapi et al. 2017; Dodangeh et al. 2020). Statistical approaches are selected to perform the selection of suitable features. For the quantification of influential factors, the Information Gain Ratio (IGR) is adopted (Towfiqul Islam et al. 2021). The greater the value of IGR, the higher the factor's influence on the target variable (Dodangeh et al. 2020). In this study, the VIF, Tolerance, correlation, and IGR values show that slope, elevation, and DTR are major factors that affects flooding in GRB. This correlation could be attributed to floodplains near rivers being more susceptible to flooding than other regions. Historical flood records also revealed that flat plains with lower slopes were the most affected region (Tehrany et al. 2015a, 2015b; Khosravi et al. 2016; Chapi et al. 2017; Bui et al. 2018a, 2018b). The IGR values are high for rainfall and NDVI, which clarifies that in GRB, rainfall and vegetation play a vital role in flooding. Around 80% of the rainfall occurs in the monsoon season from July to September, with a high magnitude of rainfall in the hill area (Panthi et al. 2015). The spatial pattern of the rainfall and NDVI is similar to the flood susceptibility pattern (Figures 2, 6, and 7). The major flooding occurs in the southern plains of GRB. The Gandak River enters the low-elevation and low-slope flat area from the higher Himalayas of more than 3000 masl to 44 masl. The rainfall and vegetation changes from low rainfall in the higher Himalayas and Trans-Himalaya to hills and plains of Terai with high rainfall, low vegetation, and loose soil (Dandekhya et al. 2017). The VIF, IGR, and correlation values in this study align with the flooding pattern. This signifies the importance of rainfall and vegetation in flood predictions. Many studies, such as (Tehrany, Jones, and Shabani 2019; Andaryani et al. 2021; Youssef et al. 2022), show that elevation is the main factor in flood mapping, similar to this study. The report by (GeoNet Connection Pvt. Ltd. 2019) shows that DTR, slope, LULC, and elevation are major factors contributing to flooding in the rivers of GRB.

5.2. Performance of the models and validation

Based on the performance of the models, it was evident that all the developed models effectively predicted the spatial distribution of flood susceptibility. The training points were the flood events in the success rate curve, while the validation points were the flood events anticipated for the prediction rate curve (Towfiqul Islam et al. 2021). Comparative analysis of models revealed nuanced differences in spatial patterns, underscoring the importance of additional metrics like MSE, RMSE, and Kappa score. Although models performed similarly in prediction evaluation, their spatial predictions differed, influencing flood susceptibility maps. Except for training AUROC, the values of all other evaluation parameters showed RF as the best-performing model. (RF) is a widely adopted tree-based machine-learning method in hydrology, renowned for mitigating overfitting issues inherent in individual decision trees while preserving predictive accuracy (Schoppa, Disse, and Bachmair 2020). Similar to (Avand et al. 2022a) and (Aldiansyah and Wardani 2023), this study shows that RF is a better approach to flood susceptibility mapping. The comparison of the result with the flood database from the Colorado Flood Observatory, Global Surface Water Dataset, and historical flood observations using Sentinel-1 SAR image showed that the results are appropriate for all models, with the results from RF and ANN being more realistic than LSTM and SVM.

5.3. Flood susceptibility mapping

According to the prediction made by all the models, areas near the river system were more prone to flooding than others. Likewise, it can be seen from Figure 6 and Figure 7 that the flood has impacted regions lying in the lower portion of the GRB to the maximum extent. The possible reason behind the extensive damage in the lower land of GRB, especially near the river portion, is that the river's velocity is high in the upper part. When the river travels from the confined high beds to the lower, it spreads and causes damage. Apart from those, studies conducted (Dandekhya et al. 2017) revealed that regular floods in this lower area have deposited large volumes of silt, which has increased the bed level of the river.

Similarly, the courses of rivers like Narayani and Rapti have shifted, threatening more livelihood in case of flood disaster. West Champaran, Gunjanagar, Mangalpur, and Madi are severely affected by floods. Likewise, the FSM maps obtained from the models revealed the areas near the Bharatpur of Nepal to be highly susceptible to flood. This result matches the flooding pattern observed in 2022 (Figure 7 and Figure 8).

Figure 8. Flooding in GRB: Settlements near Nepal-India and Nepal-China borders a)Parasi Bazaar, Nepal (https://english.ratopati.com) b) Chitwan, Nepal (https://sanjalkhabar.com) c) Champaran, India (https://www.telegraphindia.com), d) Bharatpur, Nepal (https://old.risingnepaldaily.com) e) Balmikinagar, India (https://www.indiatoday.in) f) Kagbeni, Nepal (https://www.gorkhapatraonline.com).

5.4. Transboundary flood hazard management

GRB originates from the glaciers in China, and the river flows into India via Nepal. Although people living in this region experience richer water resources, the region is vulnerable to flooding (Gupta et al. 2021). Nepal and India share the adverse effects of the floods. The GRB comprises the areas highly vulnerable to water-induced disaster, in the case of Nepal and India. For example, from 2000 to 2014, 35% of total fatalities in Nepal due to water-induced disasters occurred in GRB, which shares only 22% of the total area (Dandekhya et al. 2017). Likewise, West Champaran (Figure 8c), a major part of the study area on the Indian side, is one of the 28 most flood-affected areas in Bihar with frequent flooding(Acharya and Prakash 2019). This situation highlights the necessity of flood susceptibility mapping of GRB for proper flood transboundary flood plain management. Every year, floods affect the plain area near the Nepal-India border, which needs special attention from both governments. Similarly, glacial melting and glacial lake outburst flood in the northern side of the basin suggest including the part of the basin in China in the study. Significant improvement in the climatic knowledge and development of impact-reducing strategies in respective countries would be a critical step for generating resilient strategies for transboundary GRB basin management. A humanitarian aspect should be applied to flood inundation studies, prevention, and mitigation measures. However, for issues related to political aspects, there should be bilateral understandings and follow international norms and laws respecting each other's sovereignty (Dhungel and Pun 2009). There should be joint flood-related information sharing, joint adaptation, and mitigation strategies and technologies. An integrated approach based on scientific findings should be the primary focus for transboundary basins. Both neighboring countries should build a resilient society rather than build dams and structures that benefit certain people and harm many (Pandey et al. 2023). Early warning communication and timely actions are crucial in saving people's lives, which can be initiated by the local government level or people-to-people communication. timely communication should not be limited to Nepal and India; prompt action of the upper riparian state, People's Republic of China, can save the lives and property of the people of Nepal downstream.

A good example is the timely communication in the case of the Kagbeni flood in August 2023 (Figure 8f), where people from the upstream area communicated timely to save lives. The timely information about possible flooding based on real-time hydro-meteorological data is crucial. In the case of Nepal and India, timely information and action become crucial when opening gates in the dams, which can save lives and properties in both upstream and downstream transboundary regions (Gupta et al. 2021). Transboundary water management should be based on cooperation rather than driven by hegemony. Besides the current scenario, the future impacts due to changing climate should be studied and considered for future flood prevention and mitigation strategies. In this regard, flood susceptibility mapping plays a vital role in proper planning and policymaking. Identifying the flood susceptible area based on a scientific approach is crucial for informed decision-making (Duwal, Liu, and Pradhan 2023). Using machine learning approaches using open-sourced data is beneficial for transboundary river basins where data availability hinders scientific studies.

6. Conclusion

Due to the transboundary nature of the Gandak River Basin, the use of hydro-meteorological data from government agencies might be cumbersome. To account for this, this study used freely available open-source remote sensing data. The analysis of the reliability of these data and machine learning algorithms for flood susceptibility mapping in this study shed light on their importance, especially in the data-scarce transboundary river basins. The historical flood observation data from different sources for flood inventory preparation paved the path for robust flood susceptibility mapping where uncertainties due to the channel shifting by the river are addressed. The importance of the topographic and hydro-meteorological factors in flood generation is revealed through suitability tests such as VIF, correlation, and IGR. The highly vulnerable areas are the low-vegetated, low-plain, and low-slope areas near the river with high rainfall in the southern part of the GRB. These facts in this study signify that Slope, Elevation, and distance to river, along with vegetation, are the flow drivers. Most of Nepal and India's settlements practice agriculture and live around the floodplains near the rivers, as sediment deposited during the floods makes the land more fertile (Gupta et al. 2021). Creating buffers using vegetation could be one of the strategies to decrease the vulnerability of settlements to flooding. Among the applied machine learning algorithms, RF performed best based on different scores; however, other models also had appreciable scores. This result signifies that the high flood risk area in the GRB ranges from 4910 km² (from RF) to 9338 km² (from LSTM). The high susceptibility to flooding of around 13–24% of the area (mostly in the southern plain) describes the urgency of collaborations. This transboundary nature of flood risks emphasizes the need for bilateral cooperation in information sharing, adaptation, and mitigation strategies. Political considerations should adhere to international norms and laws, fostering resilience and societal well-being over infrastructure-centric approaches. Likewise, Flood vulnerability is increasing in the region with changing climatic conditions, land use policy, and deteriorating flood-control infrastructure. The study of erratic rainfalls and sedimentation cannot be sufficient to cater to future risks, so cooperation is essential in scientific studies and their implementations (Venkatesh Dutta 2022). Investigating the scientific method for developing probable flood risk across the basin can enhance the strategy for facing the disaster rather than a state-centric framework (Pandey et al. 2023). With technological advancement, neighboring countries should share and utilize geospatial data to develop warning systems. In this regard, flood susceptibility mapping using machine learning and Open-source data is instrumental in scientific planning and decision-making for transboundary river basins with limited data availability.

Geolocation information

The major study area lies in Nepal, with an upper portion of the basin in China and a lower part in India. It extends between the Koshi and Karnali Rivers. Gandak River is a tributary of the Ganga River. The basin lies within north latitude 29°18” and east longitude 83°85”. The tentative location in google map can be viewed through this link https://maps.app.goo.gl/2hxV9NAh2eecAswT8.

Acknowledgment

The authors thank the Research and Development Unit and Technical Section, Khwopa College of Engineering, for their constant support throughout the project.

Disclosure statement

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.