A GeoAI framework for coastal flood risk assessment: integrating remote sensing and socioeconomic data

Abstract

Bangladesh’s coastal zone is widely recognized as one of the most hazard-prone regions of the world. Within this vulnerable belt, Lakshmipur District stands for its long, highly exposed coastline and its recurrent experience of hydrometeorological hazards, including frequent floods, tidal surges, river overflows, and intense monsoon rainfall. The district’s flat topography and geographic setting further amplify its susceptibility to both seasonal and extreme flood events. Notably, two consecutive major floods in 2024 and 2025 affected approximately 600,000 and 50,000 people, respectively, with the 2024 event alone causing crop losses exceeding BDT 2.27 billion. In response to these escalating risks, this study develops a Geospatial Artificial Intelligence (GeoAI)-based flood risk mapping framework that integrates multi-source satellite Earth observation data and social datasets to map flood extents for 2024 and 2025 and to predict flood risk for Lakshmipur District. Flood extents were derived using Sentinel-1 Synthetic Aperture Radar (SAR) imagery, while flood risk prediction employed a Random Forest classifier trained on elevation, slope, distance to rivers, precipitation, population, cropland, and built-up indices. The resulting probabilistic flood risk was classified into five categories ranging from very low to very high risk, with population and cropland exposure quantified at the union level. Model validation using 2025 flood data achieved an overall accuracy of 85.9% and a Cohen’s Kappa of 0.71, demonstrating strong predictive performance. The results highlight critical flood-risk and exposure hotspots, underscoring the utility of GeoAI-based approaches for enabling timely, efficient, and actionable flood risk assessments, particularly for vulnerable communities where early risk detection through mapping and modeling is essential for proactive disaster planning and targeted mitigation for natural disaster management.

1 Introduction

Across the world, flooding is among the most common and costly natural hazards, contributing to approximately 40% of weather-related disasters and imposing estimated economic losses of approximately USD 388 billion annually (UNDDR, 2025). Climate change has significant negative impacts on coastal ecosystems, infrastructure, and human settlements (Tahir et al., 2026). Climate change and sea-level rise are expected to increase the vulnerability of people living in coastal areas and delta’s worldwide (Passeri et al., 2015). Additionally, damages from coastal flooding are expected to increase unless existing coastal protection measures are improved globally (Vousdoukas et al., 2026). The Ganges-Brahmaputra-Meghna (GBM) delta, located in Bangladesh and India, is home to millions of people who live under the regular threat of flooding and riverbank erosion, frequently displacing families from their homes, especially during the monsoon season (Curtis et al., 2018; Rahman et al., 2021; Islam and Mitra, 2025).

Bangladesh is highly susceptible to natural hazards and disasters due to its distinctive geographical, hydrological, and meteorological characteristics. The nation regularly confronts a range of natural disaster-related challenges, such as droughts, tropical cyclones, flooding, riverbank erosion, storm surges, and salinity intrusion (Rahman et al., 2024). Frequent flood events pose significant threats and inundate large portions of Bangladesh each year (Islam et al., 2026). Recurrent flood events often affect one-third of the country annually, with impacts that vary across space and time and include loss of lives, damage to property, infrastructure, and livelihoods (Islam et al., 2026; Paul, 1997). Bangladesh experiences several types of flooding, including fluvial (river) floods, pluvial (rainfall-induced) floods, flash floods, tidal floods, and storm surge-driven coastal floods driven by south-western monsoon winds (Islam and Sharabony, 2023; Shampa et al., 2025).

Flood events occur across both coastal areas and inland regions, but in coastal districts, they are often driven by the combined influence of hydrometeorological factors and dynamic estuarine processes. Coastal Bangladesh, particularly districts such as Lakshmipur, is highly susceptible to flooding due to its low-lying deltaic geomorphology, the junction and network of rivers, and monsoonal rainfall. Lakshmipur District, situated along the Meghna estuary, is particularly susceptible to monsoon flooding, tidal surges, and river overflows due to its low elevation and the juncture of multiple rivers. Flooding in Lakshmipur occurs due to the combined influence of intense monsoon river discharge from the Ganges-Brahmaputra-Meghna (GBM) river system, tidal influences, low-lying deltaic topography, and an exposed coastline (Ghosh et al., 2023). During the monsoon season, high upstream flows interact with high tides, producing extensive inundation. Additionally, tropical cyclone-induced storm surges funneling into the estuary often cause coastal flooding. Lakshmipur District experienced repeated flooding events in both 2024 and 2025, disrupting people’s livelihoods and causing extensive damage and economic losses (Dhaka Tribune, 2024a). The Flood of 2024 is shown in Figure 1.

Figure 1

In August 2024, heavy monsoon rainfall together with unusually strong tidal influence impacted approximately 600,000 people across hundreds of villages and several municipal areas, resulting in widespread inundation of homes, cropland, seedbeds, homestead assets, fish ponds, and aquaculture (Dhaka Tribune, 2024b). Subsequently, in May 2025, heavy rainfall and powerful tidal surges from the Meghna River inundated more than 100 villages in four coastal upazilas, cutting power supply, disrupting communication networks, and severely affecting the daily lives of thousands of people (Standard, 2025). Following these two flood events, large portions of agricultural land, including Aman rice seedbeds, vegetable fields, and aquaculture ponds, sustained extensive damage, along with significant infrastructure losses, further reflecting the district’s continued vulnerability to monsoonal and tidal flooding.

For flood mapping, Synthetic Aperture Radar (SAR) has become the preferred remote sensing dataset, particularly in monsoon-affected regions such as Bangladesh, where persistent cloud cover during the flood season renders conventional optical satellite imagery largely ineffective (Buchroithner and Granica, 1997; Uddin et al., 2021). Unlike passive optical sensors, which depend on reflected solar radiation and are obstructed by clouds, SAR is an active microwave system that transmits and receives its own energy, enabling it to penetrate cloud cover, operate under rainy conditions, and acquire imagery both day and night (Flores-Anderson et al., 2019). This all-weather, day-night capability is critical in Bangladesh, where flood peaks frequently coincide with the densest cloud cover during the monsoon season, making SAR data the most operationally appropriate tool for near real-time inundation monitoring (Uddin et al., 2021). Building on these properties, SAR satellites have been increasingly and effectively applied to flood detection and inundation mapping across Bangladesh in recent years (Ahmed et al., 2025; Islam et al., 2026; Martinis et al., 2015). These studies have collectively demonstrated that multi-temporal SAR analysis can capture not only the spatial extent of flooding but also its progression and recession over time, supporting both emergency response and post-flood damage assessment.

To address such flood risks, advances in GeoAI offer increasingly powerful tools for flood risk mapping. GeoAI is a fast-growing discipline that combines innovations in spatial data science, Artificial Intelligence (AI), Machine Learning (ML), Deep Learning (DL), and big geospatial data analysis and modeling (VoPham et al., 2018). In recent years, the application of GeoAI in hydrological and fluvial systems, as well as other hydrological subfields, has been rapidly advancing, with the capacity to incorporate multiple spatial–temporal scales and risk mapping (Gonzales-Inca et al., 2022). A critical moment in the evolution of GeoAI occurred during the 2017 ACM SIGSPATIAL conference, where a dedicated workshop formally articulated GeoAI as the integration of AI methods with geographic information and other geospatial data (Hu et al., 2019; Mao et al., 2017). Since then, several studies have advanced the integration of machine learning algorithms with topographic and geological characteristics to model flood susceptibility and mapping (Apriana et al., 2024; Mubeen and Saeed, 2023; Rezvani et al., 2024).

The GeoAI combined with Google Earth Engine (GEE) has become a powerful platform for large-scale flood mapping and monitoring, enabling rapid processing of Earth Observation (EO) data while integrating machine-learning approaches. Studies demonstrate how GEE’s cloud infrastructure enhances predictive performance and operational readiness: Jahani et al. (2024) showed that coupling GEE-based hydrological inputs with AI models substantially enhances flood-flow prediction accuracy; Mehmood et al. (2021) used Landsat imagery on GEE to map flood extent efficiently with minimal preprocessing; and Tian et al. (2026) introduced DeepSAR Flood Mapper, which uses Sentinel-1 SAR, the HAND index, and a multilayer perceptron model on GEE to produce global-scale flood maps with high precision. Complementing these advances, Khan et al. (2024) demonstrated a real-time GEE-machine-learning framework capable of rapidly analyzing flood impacts across large floodplains. Collectively, these contributions show that GeoAI-GEE workflows significantly enhance the speed, scalability, and reliability of flood mapping in complex hydrological environments.

Recent advances have further expanded GeoAI capabilities for flood risk beyond extent mapping. Ensemble machine learning methods, such as Random Forest, XGBoost, and hybrid deep learning architectures, have demonstrated superior performance over conventional statistical models in flood susceptibility mapping by capturing complex non-linear relationships among topographic, hydrological, and climatic variables (Mosavi et al., 2018; Costache et al., 2022). Transformer-based models and graph neural networks have also recently been applied to spatio-temporal flood forecasting, enabling dynamic risk prediction at finer temporal resolutions (Bentivoglio et al., 2022). Critically, however, most of these advances remain focused on hazard delineation rather than integrated risk assessment that couples physical flood drivers with socioeconomic exposure and vulnerability. This study directly addresses this gap.

Despite the rapid expansion of GeoAI in geospatial modeling and hazard assessment, its application in flood risk studies that integrate both social and environmental variables remains limited, particularly in complex coastal regions. Much of the existing literature continues to emphasize physical drivers of flooding, including elevation, slope, rainfall, drainage networks, sediment dynamics, tidal river morphology, and other hydrodynamic processes, without linking these factors to flood susceptibility or community-level impacts (Faisal and Hayakawa, 2022; Islam and Rahman, 2024; Rahman et al., 2019). However, in Lakshmipur, where dense rural populations, extensive rice cultivation, aquaculture activities, and recurring economic losses from seasonal inundation intersect with a highly dynamic deltaic landscape, a more integrated social-environmental approach is essential. The existing literature rarely combines multi-sensor earth observation (EO) data with social data, leaving a limited understanding of how environmental hazards translate into localized impacts at smaller administrative units, such as unions (the lowest administrative unit in Bangladesh), which typically comprise approximately 10 villages. This gap partly reflects a conceptual limitation in existing literature. Many studies conflate flood hazard with flood risk, treating inundation extent as a proxy for impact. Risk, however, is understood as the product of hazard, exposure, and vulnerability (UNDDR, 2025; Koks et al., 2015), whereas hazard refers to the physical flood event, exposure to the people and assets in flood-prone areas, and vulnerability to the susceptibility of those elements to suffer harm. Flood hazard mapping alone, without integrating who and what is exposed and how susceptible they are, systematically underestimates true risk and can misdirect adaptation investments toward physically flood-prone but sparsely populated areas while overlooking densely settled, socioeconomically fragile zones, such as rural coastal Bangladesh.

To address these gaps, this study pursues three interconnected objectives. First, it maps the flood extents for both the 2024 and 2025 flood events in Lakshmipur District using multi-temporal Sentinel-1 satellite imagery. Building on these flood extent maps, it then deploys a GeoAI-based machine learning approach within GEE to predict flood risk by integrating selected environmental, hydroclimatic, and social variables. Finally, the study quantifies population and cropland exposure within the predicted flood-risk areas to support proactive measures and reduce future losses from flooding.

2 Materials and methods

2.1 Study area

Lakshmipur District (Figure 2) is a coastal district in southeastern Bangladesh, located in the Chittagong Division and covering an estimated area of 1,456 km². Geographically, the district lies between 22°30’–23°10’N and 90°38’–91°01’E and shares borders with Chandpur District to the north, Noakhali District to the east, and the Meghna River to the south and west. The district consists of five Upazilas (the second lowest administrative unit in Bangladesh)–Lakshmipur Sadar, Raipur, Ramganj, Ramgati, and Kamalnagar–encompassing 58 unions and 547 villages. The total population of Lakshmipur District is approximately 1,937,948 across 459,344 households, predominantly relying on agriculture and services for their livelihoods (BBS, 2022).

Figure 2

The landscape of Lakshmipur District is predominantly deltaic, characterized by flat, low-lying alluvial and tidal floodplains with elevations ranging from 3 to 8 m above mean sea level. The physical geography of the region is dominated by several rivers and canals, including the Dakatia and Katakhali rivers, and is heavily influenced by the Lower Meghna River system. Lakshmipur experiences a subtropical monsoon climate marked by high humidity and substantial annual rainfall averaging 3,000 to 3,300 mm. Most of the precipitation occurs during the monsoon season (June–September). The mean annual temperature is approximately 25.7 °C, with peak temperatures commonly reaching 34.3 °C during April and May, while January is the coolest month, with average minimum temperatures of approximately 14.4 °C. Although these climatic conditions are favorable for diversified agriculture, they also render the district highly vulnerable to seasonal hazards, such as cyclones, tidal surges, and flooding. However, in recent years, tropical cyclones, tidal surges, and riverbank erosion have already impacted the livelihoods of the residents along the riverbank and in inland areas (Crawford et al., 2021; Jahan, 2023).

The western and southern parts of the district, particularly Ramgati and Kamalnagar Upazilas, experience persistent riverbank erosion (Crawford et al., 2020 and 2021). Soil consists mainly of calcareous gray floodplain sediments, supporting intensive agricultural activities. The district is known for rice, soybean, coconut, and betel nut production. Additionally, the floodplain, riverine, and estuarine environments and ecosystems offer important economic opportunities for fisheries, including Hilsa (Tenualosa ilisha), providing a significant source of both income and protein for households living in the region.

2.2 Data

The analysis draws on a suite of openly accessible geospatial datasets available through the Google Earth Engine (GEE) platform. Sentinel-1 SAR imagery was used to map flood extent for the years 2024 and 2024. Topographic variables such as elevation and slope were derived from NASA’s SRTM digital elevation model at 30-m resolution and treated as static for both study years. Hydrological context was represented using the HydroSHEDS river network, which was converted to a 30-m raster to generate Euclidean distance-to-river metrics. Seasonal monsoon precipitation was obtained from the CHIRPS v2.0 rainfall archive (5 km resolution, resampled to 30 m) from May to September of 2024 and 2025. Land use indicators included a built-up index Normalized Difference Built-up Index generated from Sentinel-2 MSI (10 m, resampled to 30 m) and a 10-m cropland extent map derived from Sentinel-2 imagery for March 2024. Population density was represented using the WorldPop 2024 gridded population dataset (100 m resolution, resampled to 30 m). For training and validation, annual flood extents were mapped using Sentinel-1 SAR (VV polarization) at 10 m resolution, comparing pre-flood and peak-flood periods for both 2024 and 2025. Together, these datasets provided consistent and high-quality inputs for modeling flood susceptibility, exposure, and spatial risk patterns across the district. We used open datasets available through the GEE platform, which are widely used for flood-related study and geospatial analysis and mapping. In Table 1, data details are provided.

2.3 Flood risk mapping

To establish a uniform analytical framework, multi-source geospatial datasets were obtained and managed using Google Earth Engine (GEE) and ArcGIS Pro. Before modeling, all explanatory variables were aligned with reprojecting to a singular coordinate system, spatial resampling to a 30-m grid cell resolution, bounding box clipping to the study area, and Min-Max normalization onto a 0 to 1 scale. Topographic and hydrological data, specifically elevation and slope extracted from the SRTM Digital Elevation Model (DEM) along with proximity to water channels derived from the HydroSHEDS stream network, were classified as static parameters due to their stability throughout the study period. In contrast, time-variant environmental indicators were incorporated to account for seasonal variations across the 2024 and 2025 monsoon periods. These dynamic factors comprised cumulative monsoon precipitation (spanning May to September) sourced from the CHIRPS archive, as well as surface imperviousness trends quantified via the Normalized Difference Built-up Index (NDBI) computed from Sentinel-2 multi-spectral bands. Agricultural landscapes were delineated using a March 2024 Sentinel-2 baseline cropland product, and demographic variations were represented using the 2024 WorldPop gridded dataset to ensure a highly dependable population baseline. Finally, historical inundation footprints for both 2024 and 2025 were mapped from dual-polarization Sentinel-1 SAR observations collected during baseline pre-flood and maximum extent peak flood phases, serving as the core empirical target arrays for machine learning training and verification.

2.3.1 Flood detection using synthetic aperture radar (SAR)

In this study, Sentinel-1 SAR data were accessed and processed using the GEE cloud computing platform. VV polarization Ground Range Detected (GRD) scenes were selected because VV polarization is particularly sensitive to the presence of surface water, producing strong contrast between flooded and non-flooded areas. GEE distributes Sentinel-1 GRD data with standard preprocessing already applied, encompassing orbital file correction, thermal noise removal, radiometric calibration, and terrain correction (Filipponi, 2019). A Lee speckle filter with a 5 × 5 kernel window was applied within the GEE environment to reduce speckle, the characteristic granular noise inherent to all SAR imagery that arises from the coherent interference of radar returns, thereby improving the visual interpretability and classification accuracy of the imagery (Islam et al., 2026).

To map the seasonal flood dynamics across the 2024 and 2025 monsoon cycles, we established a change-detection workflow centered on automated backscatter thresholding. For each study year, an initial pre-flood reference baseline was built by calculating the mean backscatter from a 3-month window preceding May. This dry season reference was then subtracted from a corresponding peak-monsoon mean composite spanning active May to August inundation window. Because smooth open water acts as a specular reflector that bounces radar signals away from the satellite’s line of sight, active flooding generates a pronounced drop in backscatter intensity (dB) relative to the rougher textures of unflooded landscapes. Rather than manually selecting an arbitrary cutoff value, we extracted the 90th percentile of this backscatter difference surface across the district to serve as an objective, data-driven threshold. This automated binarization step isolated the true anomalies, yielding a clean binary raster in which pixel values of 1 denoted standing floodwater and 0 indicated dry land. These final spatial layers provided the empirical ground-truth matrices required to train and validate our Random Forest-based prediction model. A detailed description of the methodology followed for this study is provided in Figure 3.

Figure 3

2.3.2 Flood risk prediction

Seven predictor variables were selected to represent the key environmental, hydroclimatic, and social drivers of flood risk in Lakshmipur. Elevation data were derived from the Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM). It was included as a primary topographic control, given that lower-lying areas are inherently more susceptible to inundation (Asrade et al., 2026). Slope, also calculated from the SRTM DEM, captures terrain gradient; flatter slopes impede surface drainage and promote water retention, thereby amplifying flood risk. Distance to river, computed as the Euclidean distance from the HydroSHEDS river network, reflects proximity to water channels, with areas located closer to rivers facing a greater likelihood of inundation during high-discharge events. Precipitation was represented using Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) rainfall estimates, averaged over the 2024–2025 monsoon seasons, to capture the hydroclimatic forcing associated with the observed flood events. Land surface imperviousness was characterized by calculating the Normalized Difference Built-up Index (NDBI). NDBI was derived from Sentinel-2 imagery to quantify the spatial extent of impervious surfaces, serving as a key indicator of human-altered terrain that limits soil infiltration and exacerbates runoff generation (Wang et al., 2023). It can be calculated using the Equation 1:

Additional variables include population density, drawn from the WorldPop 100 m gridded dataset, and cropland extent, derived from a Sentinel-2-based 2024 cropland dataset to delineate agricultural land at risk of flood damage. To normalize all the continuous variables, we used the following Min-Max scaling Equation 2:

This transformation rescaled all variables to a common range of 0 to 1, ensuring that no single variable exerted disproportionate influence on the model due to differences in magnitude rather than predictive importance. Following harmonization, all normalized layers were stacked into a multi-band raster, with each pixel containing a consistent vector of predictor values representing the full suite of environmental and social characteristics at that location. This multi-dimensional feature space formed the input to the Random Forest classifier and was used to train the model using the 2024–2025 SAR-derived flood labels.

Finally, using the Random Forest (RF) classifier, the flood risk probability was predicted using Equation 3. RF is preferred over other machine learning models for flood risk mapping because it handles non-linear environmental relationships, requires no data normalization, and is highly resistant to overfitting (Yousaf, 2025). The classifier consisted of 300 decision trees, implemented with the GEE platform. The prediction is expressed as:

Where N denotes the number of trees (300), Ti represents the prediction of each decision tree, and X is the vector of predictor variables at each pixel.

Based on the methodology used, the predictive flood risk model captured the localized statistical dependencies between the predictor stack and historical flood targets, allowing for spatial susceptibility projections across the study area. This optimized model was applied to the harmonized predictor environment to output a continuous spatial surface of flood probabilities. Finally, to translate these raw probabilities into actionable decision-making and visualization of the risk area, the continuous surface was segmented into five discrete risk classes (Very Low, Low, Moderate, High, and Very High) using a Quantile classification scheme using ArcGIS Pro. Splitting the class intervals precisely at the 20th, 40th, 60th, and 80th percentiles directly accommodates the heavy-tailed distributions typical of coastal hazard data, preventing extreme spatial outliers from distorting risk boundaries (Borga et al., 2011; Tehrany et al., 2019). This percentile-based slicing maximizes visual contrast and ensures that each risk zone occupies a mathematically equal geographic portion of the study area, eliminating ambiguous spatial clustering in low-lying terrains (Prasertsoong and Puttanapong, 2025). The final predictive flood risk map was compiled using sequential cartographic color schemes to establish a clear and intuitive visual hierarchy. By translating raw predictive indices into structured, standardized exposure tiers, this classification system closely aligns with contemporary disaster management frameworks that champion comprehensive spatial workflows for vulnerability mapping and risk assessment (Devi et al., 2025).

2.3.3 Model validation

The Random Forest model was trained using 6,000 stratified samples based on the 2024 flood extent map, with an equal number of flood and non-flood observations to ensure balanced learning. The validation of the model was conducted using an independent set of 3,000 samples derived from the 2025 flood extent map, representing a training-to-validation ratio of 67:33. This year-to-year validation approach enabled the model to be tested on a separate flood event, providing a more realistic assessment of its predictive performance. The validation results yielded 1,281 correctly classified non-flood samples and 1,295 correctly classified flood samples, demonstrating strong predictive capability and better performance for flood risk assessment. The predictive performance of the Random Forest classifier was evaluated using independent 2025 SAR-derived flood extent data as test observations, enabling an assessment of the model’s temporal transferability beyond the training period. Overall accuracy and Cohen’s Kappa coefficient were selected as the primary validation metrics, as both are widely used and well-established measures in flood susceptibility mapping (Tehrany et al., 2015; Liuzzo et al., 2019). These metrics were derived from a confusion matrix constructed from stratified test samples comparing model predictions against observed flood and non-flood locations. The confusion matrix is provided in Table 2.

Overall accuracy, defined as the proportion of correctly classified pixels across both classes, was calculated using Equation 4 as:

Where True Positives (TP) represent locations where the model correctly predicted flooding, while True Negatives (TN) represent areas correctly classified as non-flooded. Cohen’s Kappa coefficient, which adjusts for agreement occurring by chance, was calculated as κ = 0.71, indicating strong agreement between the predicted flood risk and SAR-derived observations. In addition, the model’s discriminative ability was assessed using Receiver Operating Characteristic (ROC) curve analysis, yielding an Area Under the Curve (AUC) value of 0.88 (Figure 4). Together, these validation results confirm that the Random Forest model achieves strong predictive performance and reliable spatial discrimination between flood-prone and non-flooded areas.

Figure 4

2.3.4 Flood risk exposure estimation

Population and cropland exposure to predicted flood risk were quantified at the union level using a pixel-based probability weighting approach. For each pixel, the predicted flood probability, P(Flood), was applied as a weight to the underlying social and agricultural values, and the resulting products were aggregated within each union using zonal statistics. Population exposure was computed as the sum of the product of flood probability and population count across all pixels within a given union (Equation 5):

Cropland exposure was calculated by multiplying the flood probability by the pixel area and the cropland fraction at each pixel, yielding an estimate of the expected agricultural land at risk, expressed in hectares (Equation 6):

These union-level estimates provide a spatially explicit basis for identifying communities and agricultural areas facing the greatest flood-related risk and serve as the foundation for the exposure hotspot analysis presented in the results.

3 Results

The model is trained using flood extent data from 2024 and validated against independent observations from 2025. Using this validated framework, a predictive flood risk map is generated, classified into five levels from very low to very high, and integrated with union-level population (WorldPop) and cropland (Sentinel-2) data to quantify exposure. In 2024, flooding covered approximately 550 km², or 38% of the district (Table 3). In 2025, flooding decreased to approximately 470 km², exposing an estimated 5% of the district’s population (60,500 people) and 22% of cropland (10,380 ha). The reduction in flood extent from 2024 to 2025 does not necessarily indicate diminishing risk; rather, it reflects interannual variability in monsoon intensity and highlights the unpredictable nature of flood cycles in deltaic coastal systems increasingly influenced by climate variability (Kundzewicz et al., 2019). The flood extent for 2024 and 2025 is shown in Figure 5.

Figure 5

Population and cropland exposure estimated in the model for the predicted floor risk map show that tens of thousands of people and thousands of hectares of agricultural land are likely to be impacted under high flood risk. The predicted flood risk model identifies the southwestern part of Lakshmipur District as the highest flood-risk zone because multiple factors that trigger flood events converge there. This spatial concentration of risk in the southwest aligns with broader findings on compound flood risk in delta-coastal transition zones, where riverine, tidal, and rainfall-driven processes interact to amplify inundation likelihood (Han and Tahvildari, 2024). This area consistently showed strong flood signals in the 2024–2025 Sentinel-1 data and contains some of the district’s lowest and flattest terrain, allowing water to accumulate easily. More importantly, its proximity to major Meghna River channels increases exposure to overbank flooding, while high seasonal rainfall and extensive cropland further promote water retention. The predicted 5-class flood risk map (Figure 6) shows the highest risk near major rivers and low-lying central areas of Lakshmipur District, while the high-elevated northern areas exhibit lower risk.

Figure 6

The predicted flood risk map indicates that flood susceptibility is distributed throughout Lakshmipur District, with varying levels of risk ranging from very low to very high. Based on our analysis, approximately 359,000 people and 27,000 hectares of cropland are potentially vulnerable to future flood events. These critical zones are primarily concentrated within coastal and river-adjacent unions, where dense populations and intensive agricultural activities intersect with the most severe flood routes. The study results indicate how localized social and ecological exposure meets with physical hazards, such as floods, underscoring the urgent need for immediate and targeted measures at the local or union level for risk reduction and climate mitigation strategies in coastal Bangladesh.

In Figures 7A,B, population and cropland (in ha) exposures are shown as predicted flood risks areas.

Figure 7

To determine the final top 10 priority zones, we aggregated and sorted the joint metrics obtained from the predicted flood risk surface area for the study area by union. Because the individual rankings for population and cropland were determined globally across the entire study area dataset, some individual component ranks within the top tier naturally exceed 10. This pattern shows that a union does not need to be in the top 10 for every single metric to be considered highly vulnerable overall; a strong combined exposure profile is what drives its high priority status. Our analysis shows that flood exposure varies drastically throughout the district. At the local level, exposed cropland reaches nearly 2,700 hectares per union, while the exposed population ranges from less than one person to over 21,000 residents. The highest population exposure figures are clustered in North Char Bangshi (exceeding 21,000) and Char Ramani Mohan (approaching 20,000) (Table 4). We observed a very similar spatial concentration for agricultural assets. The largest expanses of exposed cropland are found in Char Ramani Mohan (nearly 2,700 ha) and North Char Bangshi (nearly 1900 ha). The prominence of unions such as Char Ramani Mohan, North Char Bangshi, and Char Kadira is among the top risk localities. These zones consist of recently accreted, low-lying riverine islands marked by a heavy reliance on agriculture and sparse infrastructure, which leaves their populations disproportionately exposed and with limited adaptive capacity.

Exposure values reflect the scenario and inputs for the predicted flood risk and should be seen as the potentially affected number of people and cropland, not as probabilities of inundation.

4 Discussion

In managing flood risks, the Government of Bangladesh has implemented a range of strategies to reduce flood-related losses and damages, incorporating both structural and non-structural measures (Ansari et al., 2022; Iqbal and Kutubuddin, 2021). Nonetheless, rapid population growth, unplanned urban expansion, land-use transformation, and intensifying hydrometeorological extremes continue to challenge sustainable flood risk reduction, particularly in coastal Bangladesh. Coastal districts face compound risks from riverine flooding, cyclones, storm surges, and sea-level rise, which are projected to intensify under climate change (Aerts et al., 2018; Dodman et al., 2022).

Despite national planning instruments, such as the Bangladesh Delta Plan 2,100, research using cutting-edge approaches (e.g., GeoAI, Hybrid Modeling) in flood risk mapping is still limited. Many flood risk assessments rely on static hazard zonation and historical flood frequency, limiting their ability to capture dynamic environmental and socioeconomic changes (Merz et al., 2020). Moreover, conventional approaches often insufficiently integrate exposure and vulnerability indicators, such as population exposure and damage to cropland, into predictive flood risk modeling (Koks et al., 2015; Winsemius et al., 2016). These limitations show the need for transferable, efficient, data-driven frameworks that combine hazard, exposure, and vulnerability components for natural disaster modeling and mapping.

Prior flood-mapping studies in Bangladesh have largely focused on hazard delineation or static susceptibility without integrating socioeconomic exposure (Chowdhury et al., 2025; Rudra and Sarkar, 2023). The present study directly addresses this gap by combining Sentinel-1-derived flood extents with population and cropland exposure at the union level. The model produces risk estimates that are both spatially explicit and socioeconomically grounded. The finding that approximately 156,000 residents live in high- to very-high flood-risk zones, roughly one-third of the total exposed population, underscores that hazard extent alone substantially underestimates true risk when development indicators are excluded. Bangladesh-specific modeling research has limited integration of existing social and environmental perspectives as well as future climate change impacts and scenarios, constraining its utility for long-term adaptation planning. In recent years, rapid population growth, evolving socioeconomic dynamics, and increasingly complex hydrometeorological threats associated with natural hazards and disasters continue to pose significant challenges to effective and sustainable flood management and disaster risk reduction strategies, especially for coastal Bangladesh. Due to technological advancement in geospatial modeling for predicting flood risks and vulnerability, GeoAI-integrated geospatial tools and techniques offer enormous potential for delineating flood extent and damage assessment using EO and socioeconomic data for emergency management and proactive policy measures for response to flood events.

Predictive modeling is particularly valuable in data-scarce deltaic environments, such as Bangladesh, where flood behavior is driven by complex, non-linear interactions that static observational approaches cannot adequately capture (Kundzewicz et al., 2019). A direct overlay of observed flood extents with population and cropland data would capture exposure for those specific events but cannot generalize beyond observed conditions. The predictive framework, by contrast, learns the spatial relationships among terrain, proximity to river channels, land cover, and flood occurrence, enabling risk estimation across the full district, including unions that were not inundated during the training period but share similar physical characteristics with those that were. This distinction between event-specific exposure mapping and probabilistic risk modeling is critical for proactive planning rather than retrospective damage assessment.

The model’s validation against 2024 and 2025 flood events demonstrates that GeoAI frameworks trained on recent Earth observation data can generate meaningful predictive flood risk outputs for near-term planning horizons. The GeoAI framework contributes beyond conventional GIS overlay in three specific ways: First, the Random Forest classifier captures non-linear interactions among flood-driving variables that rule-based overlay methods cannot detect. Second, it produces a continuous risk surface that can be updated as new Sentinel-1 observations become available without manual re-digitization. Finally, validation against an independent year (2025) provides a quantitative performance benchmark absent from most static flood susceptibility maps in the region. This is particularly relevant for coastal Bangladesh, where the Bangladesh Delta Plan 2,100 calls for adaptive, evidence-based flood management but lacks union-level probabilistic risk tools. The identification of char-dominated unions as the highest combined risk zones further reinforces findings from the broader coastal vulnerability literature, which consistently links land tenure insecurity, agricultural dependence, and low elevation to heightened disaster risk in deltaic South Asia (Hill et al., 2019; Becker et al., 2024). Taken together, these results demonstrate that integrating remote sensing-derived hazard data with socioeconomic exposure indicators within a GeoAI framework meaningfully advances flood risk assessment beyond conventional static approaches.

Despite its novel contributions, this study has several limitations. First, the model was trained exclusively on Sentinel-1 SAR-derived flood extent data from 2024 and 2025, which represents a limited temporal window and may not capture the full range of flood variability in the region. Second, while the Random Forest classifier performed well in validation, its predictions are inherently constrained by the quality and resolution of input variables (Fox et al., 2017). Finer-scale socioeconomic indicators, such as household income, housing quality, and access to early warning systems, were not incorporated, limiting the model’s ability to capture within-union vulnerability differences. Third, union-level averaging may obscure intra-union heterogeneity, where localized very high-risk settlements near unprotected embankments could be masked by broader moderate-risk classifications, limiting utility for sub-union humanitarian targeting. Finally, the framework does not explicitly account for future climate change scenarios, such as projected sea-level rise or shifts in monsoon intensity, which are critical for long-term flood risk planning in deltaic environments. Future studies should address these gaps by incorporating multi-temporal flood records, household-level vulnerability indicators, and climate projection data to enhance the predictive and adaptive value of GeoAI-based flood risk frameworks.

5 Conclusion

This study demonstrates how a GeoAI-based geospatial modeling approach can provide a better understanding of flood risk in one of Bangladesh’s disaster-vulnerable coastal districts. By incorporating multiple datasets (e.g., remote sensing data, environmental variables, and social data) to map recent flood events (2024 and 2025) in Lakshmipur District, this study provides a critical foundation for anticipating future flood events. The Random Forest model used in the analysis matched the 2025 flood extent and showed that combining SAR-derived flood mapping with machine learning techniques can reveal the influence of terrain, hydrology, land use, and population exposure to flooding. The study results produce risk maps that highlight where people and croplands face the greatest threats, offering a detailed flood risk map at the union level of the study area.

The study results can support local and regional disaster risk reduction, emergency planning, and decision-making by identifying where resources and interventions are most urgently needed. The framework developed here also has strong potential as a decision support tool for flood management in Bangladesh and beyond. Expanding this approach to neighboring districts or other coastal regions and incorporating additional socio-economic and environmental factors could help create a transferable model for areas facing similar challenges. Ultimately, this study underscores how GeoAI can strengthen early warning systems and improve resilience for communities living on the front lines of recurring, intensifying flood hazards.

References

• 1

  AertsJ. C.BotzenW. J.ClarkeK. C.CutterS. L.HallJ. W.MerzB.et al. (2018). Integrating human behaviour dynamics into flood disaster risk assessment. Nat. Clim. Chang.8, 193–199. doi: 10.1038/s41558-018-0085-1

  • CrossRef
  • Google Scholar

• 2

  AhmedR. U.RashidM. M.HabibE. H. (2025). Flood extent mapping from SAR images by a dynamic threshold: a case study of 2022 compound flood in northeastern Bangladesh. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.19, 3488–3501. doi: 10.1109/jstars.2025.3647843

  • CrossRef
  • Google Scholar

• 3

  AnsariM. S.WarnerJ.SukhwaniV.ShawR. (2022). Implications of flood risk reduction interventions on community resilience: an assessment of community perception in Bangladesh. Climate10:20. doi: 10.3390/cli10020020

  • CrossRef
  • Google Scholar

• 4

  AprianaA.Al AnsoryA. R.AgustinaT.AmaliaI.RefrizonR. (2024). Mapping flood-prone areas using GIS through as geo-artificial intelligence (geo-Ai) approach in Bengkulu City. J. Technomaterial Phys.6, 56–61. doi: 10.32734/jotp.v6i1.16006

  • CrossRef
  • Google Scholar

• 5

  AsradeT. M.AbebeS. A.TadesseK. B.KerebihM. S.MesheshaT. M. (2026). Flood susceptibility assessment using three machine learning techniques and comparison of their performance. Sci. Rep.16, 8099–8099. doi: 10.1038/s41598-026-38391-0,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 6

  BBS (2022). Population and Housing Census 2022. Dhaka, Bangladesh: Government of the People’s Republic of Bangladesh.

  • Google Scholar

• 7

  BeckerM.SeegerK.PaszkowskiA.MarcosM.PapaF.AlmarR.et al. (2024). Coastal flooding in Asian megadeltas: recent advances, persistent challenges, and call for actions amidst local and global changes. Rev. Geophys.62:846. doi: 10.1029/2024RG000846

  • CrossRef
  • Google Scholar

• 8

  BentivoglioR.IsufiE.JonkmanS. N.TaorminaR. (2022). Deep learning methods for flood mapping: a review of existing applications and future research directions. Hydrol. Earth Syst. Sci. Discuss.26, 1–50. doi: 10.5194/hess-26-4345-2022

  • CrossRef
  • Google Scholar

• 9

  BorgaM.AnagnostouE. N.BlöschlG.CreutinJ. D. (2011). Flash flood forecasting, warning and risk management: the HYDRATE project. Environ. Sci. Pol.14, 834–844. doi: 10.1016/j.envsci.2011.05.017

  • CrossRef
  • Google Scholar

• 10

  BuchroithnerM. F.GranicaK. (1997). Applications of imaging radar in hydro-geological disaster management: a review. Remote Sens. Rev.16, 1–134.

  • Google Scholar

• 11

  ChowdhuryM. E.IslamA. S.ZzamanR. U.KhademS. (2025). A machine learning-based approach for flash flood susceptibility mapping considering rainfall extremes in the northeast region of Bangladesh. Adv. Space Res.75, 1990–2017. doi: 10.1016/j.asr.2024.10.047

  • CrossRef
  • Google Scholar

• 12

  CostacheR.ArabameriA.CostacheI.CrăciunA.PhamB. T. (2022). New machine learning ensemble for flood susceptibility estimation. Water Resour. Manag.36, 4765–4783. doi: 10.1007/s11269-022-03276-0

  • CrossRef
  • Google Scholar

• 13

  CrawfordT. W.IslamM. S.RahmanM. K.PaulB. K.CurtisS.MiahM. G.et al. (2020). Coastal erosion and human perceptions of revetment protection in the lower Meghna estuary of Bangladesh. Remote Sens.12:3108. doi: 10.3390/rs12183108

  • CrossRef
  • Google Scholar

• 14

  CrawfordT. W.RahmanM. K.MiahM. G.IslamM. R.PaulB. K.CurtisS.et al. (2021). Coupled adaptive cycles of shoreline change and households in deltaic Bangladesh: analysis of a 30-year shoreline change record and recent population impacts. Ann. Am. Assoc. Geogr.111, 1002–1024. doi: 10.1080/24694452.2020.1799746

  • CrossRef
  • Google Scholar

• 15

  CurtisS.CrawfordT.RahmanM.PaulB.MiahM. G.IslamM. R.et al. (2018). A hydroclimatological analysis of precipitation in the Ganges–Brahmaputra–Meghna River basin. Water10:1359. doi: 10.3390/w10101359

  • CrossRef
  • Google Scholar

• 16

  DeviK.ReddyC. C.RahulK.KhuntiaJ. R.DasB. S. (2025). A holistic methodology for evaluating flood vulnerability, generating flood risk map and conducting detailed flood inundation assessment. Sci. Rep.15:28253. doi: 10.1038/s41598-025-13025-z,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17

  Dhaka Tribune. (2024a). Lakshmipur Reels under Severe Flooding. Available online at: https://www.dhakatribune.com/bangladesh/nation/355786/600-000-stranded-as-hundreds-of-villages-flooded (Accessed May 10, 2026).

  • Google Scholar

• 18

  Dhaka Tribune. (2024b). Lakshmipur Reels under Severe Flooding. Available online at: https://www.dhakatribune.com/bangladesh/nation/355786/600-000-stranded-as-hundreds-of-villages-flooded.

  • Google Scholar

• 19

  DodmanD.HaywardB.PellingM.BrotoV. C.ChowW.ChuE.et al. (2022). Cities, Settlements and Key Infrastructure. In: Climate Change 2022: Impacts, Adaptation, and Vulnerability. Contribution of Working Group II to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change.

  • Google Scholar

• 20

  FaisalB. R.HayakawaY. S. (2022). Geomorphological processes and their connectivity in hillslope, fluvial, and coastal areas in Bangladesh: a review. Prog Earth Planet Sci9:41. doi: 10.1186/s40645-022-00500-8

  • CrossRef
  • Google Scholar

• 21

  FilipponiF. (2019). Sentinel-1 GRD preprocessing workflow. Proceedings18:11. doi: 10.3390/ECRS-3-06201

  • CrossRef
  • Google Scholar

• 22

  Flores-AndersonA. I.HerndonK. E.ThapaR. B.CherringtonE. (2019). The Synthetic Aperture Radar (SAR) Handbook: Comprehensive Methodologies for Forest Monitoring and Biomass Estimation. Washington, DC: NASA.

  • Google Scholar

• 23

  FoxE. W.HillR. A.LeibowitzS. G.OlsenA. R.ThornbrughD. J.WeberM. H. (2017). Assessing the accuracy and stability of variable selection methods for random forest modeling in ecology. Environ. Monit. Assess.189:316. doi: 10.1007/s10661-017-6025-0,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 24

  GhoshS.RoyS.IslamA.ShitP. K.DattaD. K.IslamM. S.et al. (2023). “Floods of Ganga-Brahmaputra-Meghna Delta in Context,” in Floods in the Ganga–Brahmaputra–Meghna Delta, ed. GhoshS. (Cham: Springer International Publishing), 1–17.

  • Google Scholar

• 25

  Gonzales-IncaC.CalleM.CroghanD.Torabi HaghighiA.MarttilaH.SilanderJ.et al. (2022). Geospatial artificial intelligence (GeoAI) in the integrated hydrological and fluvial systems modeling: review of current applications and trends. Water14:2211. doi: 10.3390/w14142211

  • CrossRef
  • Google Scholar

• 26

  HanS.TahvildariN. (2024). Compound flooding hazards due to storm surge and pluvial flow in a low-gradient coastal region. Water Resour. Res.60:37014. doi: 10.1029/2023wr037014

  • CrossRef
  • Google Scholar

• 27

  HillC.DunnF.HaqueA.Amoako-JohnsonF.NichollsR. J.RajuP. V.et al. (2019). “Hotspots of Present and Future Risk within Deltas: Hazards, Exposure and Vulnerability,” in Deltas in the Anthropocene, eds. NichollsR. J.AdgerW. N.HuttonC. W.HansonS. E. (Cham: Springer International Publishing), 127–151.

  • Google Scholar

• 28

  HuY.GaoS.LungaD.LiW.NewsamS.BhaduriB. (2019). GeoAI at ACM SIGSPATIAL: progress, challenges, and future directions. Sigspatial Spec.11, 5–15. doi: 10.1145/3377000.337700

  • CrossRef
  • Google Scholar

• 29

  IqbalM.KutubuddinW. (2021). Bangladesh Delta plan 2100: charting a course for sustainable ocean governance and maritime development. Bangladesh Marit. J. Spec.1, 59–76.

  • Google Scholar

• 30

  IslamM. S.MitraJ. R. (2025). Quantification of historical riverbank erosion and population displacement using satellite earth observations and gridded population data. Earth Syst. Environ.9, 375–388. doi: 10.1007/s41748-024-00460-7

  • CrossRef
  • Google Scholar

• 31

  IslamM. S.MitraJ. R.RahmanM. K. (2026). Quantifying flash flood inundation and assessing damage using satellite earth observations: the case of 2022 flash flood in Bangladesh. Geomat. Nat. Hazards Risk17:2614729. doi: 10.1080/19475705.2026.2614729

  • CrossRef
  • Google Scholar

• 32

  IslamM. K.RahmanM. M. (2024). Establishing morpho-dynamic baseline for flow-sediment management of a tidal river in the Ganges–Brahmaputra–Meghna Delta system through field measurement. Environ. Fluid Mech.24, 265–286. doi: 10.1007/s10652-024-09985-x

  • CrossRef
  • Google Scholar

• 33

  IslamM. S.SharabonyA. (2023). “Characteristics of Flood in the Meghna River basin within Bangladesh,” in Floods in the Ganga–Brahmaputra–Meghna Delta, ed. IslamM. S. (Berlin: Springer), 423–447.

  • Google Scholar

• 34

  JahanI. (2023). Mapping Landcover Change and Population Displacement of Lakshmipur District. Bangladesh due to Riverbank Erosion From 2001–2021: A Geospatial Approach

  • Google Scholar

• 35

  JahaniM.DastoraniM. T.RashkiA. (2024). Prediction of flood flows based on the combined solution of Google earth engine data and artificial intelligence models. Iran. J. Rainwater Catchment Syst.12, 51–53.

  • Google Scholar

• 36

  KhanN. S.RoyS. K.TalukdarS.BillahM.IqbalA.ZzamanR. U.et al. (2024). Empowering real-time flood impact assessment through the integration of machine learning and Google earth engine: a comprehensive approach. Environ. Sci. Pollut. Res.31, 53877–53892. doi: 10.1007/s11356-024-33090-7,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37

  KoksE. E.JongmanB.HusbyT. G.BotzenW. J. (2015). Combining hazard, exposure and social vulnerability to provide lessons for flood risk management. Environ. Sci. Pol.47, 42–52. doi: 10.1016/j.envsci.2014.10.013

  • CrossRef
  • Google Scholar

• 38

  KundzewiczZ. W.SzwedM.PińskwarI. (2019). Climate variability and floods—a global review. Water11:1399. doi: 10.3390/w11071399

  • CrossRef
  • Google Scholar

• 39

  LiuzzoL.SammartanoV.FreniG. (2019). Comparison between different distributed methods for flood susceptibility mapping. Water Resour. Manag.33, 3155–3173. doi: 10.1007/s11269-019-02293-w

  • CrossRef
  • Google Scholar

• 40

  MaoH.HuY.KarB.GaoS.McKenzieG. (2017). GeoAI 2017 workshop report: the 1st ACM SIGSPATIAL international workshop on GeoAI:@ AI and deep learning for geographic knowledge discovery: Redondo Beach, CA, USA-November 7, 2016. ACM Sigspatial Spec.9:25. doi: 10.1145/3178392.317840

  • CrossRef
  • Google Scholar

• 41

  MartinisS.KuenzerC.WendlederA.HuthJ.TweleA.RothA.et al. (2015). Comparing four operational SAR-based water and flood detection approaches. Int. J. Remote Sens.36, 3519–3543. doi: 10.1080/01431161.2015.1060647

  • CrossRef
  • Google Scholar

• 42

  MehmoodH.ConwayC.PereraD. (2021). Mapping of flood areas using landsat with Google earth engine cloud platform. Atmos.12:866. doi: 10.3390/atmos12070866

  • CrossRef
  • Google Scholar

• 43

  MerzB.KuhlickeC.KunzM.PittoreM.BabeykoA.BreschD. N.et al. (2020). Impact forecasting to support emergency management of natural hazards. Rev. Geophys.58:704. doi: 10.1029/2020rg000704

  • CrossRef
  • Google Scholar

• 44

  MosaviA.OzturkP.ChauK. W. (2018). Flood prediction using machine learning models: literature review. Water10:1536. doi: 10.3390/w10111536

  • CrossRef
  • Google Scholar

• 45

  MubeenA.SaeedS. (2023). GeoAI-driven flood susceptibility mapping and exposure assessment in Pakistan using multi-source geospatial big data and deep learning models. Front. Comput. Spatial Intell.1, 87–97.

  • Google Scholar

• 46

  PasseriD. L.HagenS. C.MedeirosS. C.BilskieM. V.AlizadK.WangD. (2015). The dynamic effects of sea level rise on low-gradient coastal landscapes: a review. Earth's Future3, 159–181. doi: 10.1002/2015ef000298

  • CrossRef
  • Google Scholar

• 47

  PaulB. K. (1997). Flood research in Bangladesh in retrospect and prospect: a review. Geoforum28, 121–131. doi: 10.1016/s0016-7185(97)00004-3

  • CrossRef
  • Google Scholar

• 48

  PrasertsoongN.PuttanapongN. (2025). An integrated framework for satellite-based flood mapping and socioeconomic risk analysis: a case of Thailand. Prog. Disaster Sci.25:100393. doi: 10.1016/j.pdisas.2024.100393

  • CrossRef
  • Google Scholar

• 49

  RahmanM. M.BelalM. E. I.HossenM. A.TabassumN. H.MehzabinJ.MumuM. N. S.et al. (2024). Assessing the climate induced livelihood vulnerability of coastal people using sustainable livelihood framework: a study in south-Central Bangladesh. Soc. Sci.13:638. doi: 10.3390/socsci13120638

  • CrossRef
  • Google Scholar

• 50

  RahmanM. K.CrawfordT. W.PaulB. K.Sariful IslamM.CurtisS.Giashuddin MiahM.et al. (2021). Riverbank Erosions, Coping Strategies, and Resilience thinking of the Lower-Meghna River Basin community, Bangladesh. In: AlamG. M. M.Erdiaw-KwasieM. O.NagyG. J.FilhoW. L.Climate Vulnerability and Resilience in the Global South: Human Adaptations for Sustainable Futures. (pp. 259–278). Berlin: Springer.

  • Google Scholar

• 51

  RahmanM.NingshengC.IslamM. M.DewanA.IqbalJ.WashakhR. M. A.et al. (2019). Flood susceptibility assessment in Bangladesh using machine learning and multi-criteria decision analysis: M. Rahman et al. Earth Syst. Environ.3, 585–601. doi: 10.1007/s41748-019-00123-y

  • CrossRef
  • Google Scholar

• 52

  RezvaniS. M.SilvaM. J. F.d AlmeidaN. M. (2024). Mapping geospatial AI flood risk in national road networks. ISPRS Int. J. Geo Inf.13:323. doi: 10.3390/ijgi13090323

  • CrossRef
  • Google Scholar

• 53

  RudraR. R.SarkarS. K. (2023). Artificial neural network for flood susceptibility mapping in Bangladesh. Heliyon9:e16459. doi: 10.1016/j.heliyon.2023.e16459,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 54

  ShampaS.NasirN. N.WineyM. M.DeyS.ZahidS. T.TasnimZ.et al. (2025). Integration of remote sensing and machine learning approaches for operational flood monitoring along the coastlines of Bangladesh under extreme weather events. Water17:2189. doi: 10.3390/w17152189

  • CrossRef
  • Google Scholar

• 55

  StandardT. B. (2025). Meghna tidal surge floods over 100 villages as incessant daylong rain batters Lakshmipur. Available online at: https://www.tbsnews.net/bangladesh/meghna-tidal-surge-floods-over-100-villages-incessant-daylong-rain-batters-lakshmipur.

  • Google Scholar

• 56

  TahirH.DinA. H. M.HusseinT. S. (2026). Future coastal inundation risk map for Iraq by the application of GIS and remote sensing. Earth7:8. doi: 10.3390/earth7010008

  • CrossRef
  • Google Scholar

• 57

  TehranyM. S.JonesS.ShabaniF. (2019). Identifying the essential flood conditioning factors for flood prone area mapping using machine learning techniques. Catena175, 174–192. doi: 10.1016/j.catena.2018.12.011

  • CrossRef
  • Google Scholar

• 58

  TehranyM. S.PradhanB.MansorS.AhmadN. (2015). Flood susceptibility assessment using GIS-based support vector machine model with different kernel types. Catena125, 91–101. doi: 10.1016/j.catena.2014.10.017

  • CrossRef
  • Google Scholar

• 59

  TianD.WangL.LiuH.CohenS.ShuS. (2026). DeepSAR flood mapper: global flood mapping on google earth engine cloud platform using MLP deep learning model with Sentinel-1 SAR imagery and HAND topographic data. GISci. Remote Sens.63:2612306. doi: 10.1080/15481603.2025.2612306

  • CrossRef
  • Google Scholar

• 60

  UddinK.MatinM. A.ThapaR. B. (2021). “Rapid Flood Mapping Using Multi-Temporal Sar Images: An Example from Bangladesh,” in Earth Observation Science and Applications for Risk Reduction and Enhanced Resilience in Hindu Kush Himalaya Region: A Decade of Experience from SERVIR, ed. UddinK. (Berlin: Springer), 201–210.

  • Google Scholar

• 61

  UNDDR. (2025). United Nations Office for Disaster Risk Reduction: GAR 2025 Hazards: Floods. Available online at: https://www.undrr.org/gar/gar2025/hazard-exploration/floods.

  • Google Scholar

• 62

  VoPhamT.HartJ. E.LadenF.ChiangY.-Y. (2018). Emerging trends in geospatial artificial intelligence (geoAI): potential applications for environmental epidemiology. Environ. Health17:40. doi: 10.1186/s12940-018-0386-x,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 63

  VousdoukasM. I.PaprotnyD.MentaschiL.MonioudiI. N.FeyenL. (2026). Coastal flood impacts and lost ecosystem services along Europe’s outermost regions and overseas countries and territories. Nat. Commun.17:188. doi: 10.1038/s41467-025-66391-7,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 64

  WangW. J.KimD.HanH.KimK. T.KimS.KimH. S. (2023). Flood risk assessment using an indicator based approach combined with flood risk maps and grid data. J. Hydrol.627:130396. doi: 10.1016/j.jhydrol.2023.130396

  • CrossRef
  • Google Scholar

• 65

  WinsemiusH. C.AertsJ. C.Van BeekL. P.BierkensM. F.BouwmanA.JongmanB.et al. (2016). Global drivers of future river flood risk. Nat. Clim. Chang.6, 381–385. doi: 10.1038/nclimate2893

  • CrossRef
  • Google Scholar

• 66

  YousafA. (2025). Random Forest and CNN-Based Hybrid Modelling Approach for Susceptibility of Flood, Master’s Thesis. Portugal: Universidade NOVA de Lisboa.

  • Google Scholar