Integrated flood and multi-hazard susceptibility mapping in Egypt’s Red Sea Mountains using AHP–machine learning, environmental sensitivity indices, and scenario-based restoration frameworks

Abstract

Flash floods represent one of the most destructive hazards in arid and semi-arid regions, causing severe damage to infrastructure, livelihoods, and ecosystems. Their assessment is often constrained by limited historical flood records and rapidly changing land-use dynamics. This study develops an integrated and explainable framework for flash flood susceptibility mapping (FSM) in Egypt’s Red Sea Mountains, a coastal zone undergoing rapid urban and tourism expansion. Multi-temporal Sentinel-1 Synthetic Aperture Radar (SAR) data were processed in Google Earth Engine using Otsu thresholding to generate dynamic flood inventories. These inventories were combined with eleven hydro-topographic and geological predictors within a Multi-Criteria Decision Analysis (MCDA) framework using the Analytic Hierarchy Process (AHP), and further enhanced by three machine learning (ML) classifiers: Random Forest (RF), Extreme Gradient Boosting (XGB), and Gradient Boosting Machine (GBM). Model evaluation demonstrated strong predictive skill, with the hybrid AHP–RF model achieving the highest accuracy (AUC = 0.95; overall accuracy = 95%). Shapley Additive exPlanations (SHAP) quantified predictor importance, confirming elevation, runoff volume, and drainage density as dominant drivers of flood susceptibility. Beyond single-hazard mapping, the study introduced an Environmental Sensitivity and Desertification Index (ESDI) and integrated it with FSM to produce a multi-hazard susceptibility map, revealing compound high-risk zones in coastal sabkhas and intensively cultivated floodplains. Scenario-based analyses under RCP 4.5/8.5 and SSP pathways projected significant expansion of high-risk zones under intensified climate forcing and unsustainable socio-economic trajectories. By aligning scenario outputs with the Food and Agriculture Organization) FAO (Standards of Practice to Guide Ecosystem Restoration (2025), the study bridges scientific diagnostics with actionable resilience planning. The integrated framework demonstrates that coupling AHP with ML not only improves predictive accuracy but also enhances interpretability and policy relevance. The outcomes provide critical evidence for disaster risk reduction, land-use management, and ecosystem restoration, offering a transferable model for climate-resilient hazard management in arid coastal environments across Africa and beyond.

1 Introduction

Flash floods represent one of the most hazardous natural phenomena, particularly in mountainous arid coastal regions, where they cause substantial loss of life, extensive infrastructure damage, and widespread disruption to economic and social systems, including adverse impacts on major tourism cities (Khan et al., 2020; Yin et al., 2023; Jonkman et al., 2024). These regions are highly susceptible to flash flooding due to rugged terrain, pronounced elevation gradients, and short-duration, high-intensity runoff episodes (Arnous et al., 2011). Forecasting flash flood events is especially critical due to high population densities and the concentration of essential infrastructure (Othman et al., 2023).

Flooding susceptibility is influenced by multiple interacting physical and environmental factors, including topography, geomorphology, drainage network characteristics, lithology, and regional climatic patterns (Mashaly and Ghoneim, 2018; Abdrabo et al., 2020; Mohamed et al., 2021; Ali et al., 2023; Abdelkareem and Mansour, 2023; Abdelgawad et al., 2024; Khattab et al., 2025; Bryndal et al., 2017; Feloni et al., 2022; Kotb et al., 2024). Despite the importance of comprehensive flood hazard assessment, assembling complete flood inventories remains challenging in many regions due to complex terrain, high operational costs, and accessibility constraints, especially in developing countries (Manandhar et al., 2023). Consequently, advanced geospatial technologies, particularly Remote Sensing (RS) and Geographic Information Systems (GIS), have become indispensable for delineating and analyzing flood-prone areas, especially in landscapes with rugged relief (El Bastawesy et al., 2019; Abuzied and Mansour, 2019; Ding et al., 2021; El-Bagoury et al., 2025). In this regard, dynamic flood inventories derived from Sentinel-1 imagery have proven effective in supporting flood mapping in data-scarce, hazard-prone arid mountainous coastal environments (Feizbahr et al., 2025).

Although several previous studies, including Youssef et al. (2022), El-Rayes et al. (2023), and Rashwan et al. (2023), have assessed flood susceptibility in the Red Sea region using GIS-based MCDA or machine-learning approaches, these investigations relied on static flood inventories and single-hazard perspectives. The present study advances beyond this work by employing multi-temporal Sentinel-1 SAR data (2016–2025) processed in Google Earth Engine to generate dynamic flood inventories, integrating AHP-derived expert weights with ensemble ML classifiers (RF, XGB, GBM), and applying SHAP-based explainable AI to quantify predictor importance. In addition, the framework extends to multi-hazard mapping by combining flood susceptibility with the Environmental Sensitivity and Desertification Index (ESDI) and incorporates RCP/SSP scenario analyses aligned with FAO ecosystem restoration standards (2025). This integrated approach provides a comprehensive and transferable methodology for resilience planning in arid coastal environments.

Recent developments in GIS-based modeling have shifted from conventional estimation techniques toward diverse Artificial Intelligence (AI) approaches. These methods are generally categorized into Multi-Criteria Decision Analysis (MCDA) and ML algorithms. While MCDA approaches are relatively straightforward, they may struggle with complex, high-dimensional datasets, particularly in regions with limited data availability (Kumar, 2025; Lu et al., 2024; Sahoo and Goswami, 2023; Wu et al., 2022). Traditional MCDA methods, including knowledge-driven (KD) approaches and fuzzy logic (FL), often suffer from lower predictive accuracy and slower computation (Sahoo and Goswami, 2023; Abedi Gheshlaghi and Feizizadeh, 2021; Beaudrie et al., 2020; Megahed et al., 2025). For example, Dieno and Pathan (2025) employed AHP with 3D flood simulations and multi-temporal satellite imagery to develop a flood susceptibility model for Sonitpur City, India. Similarly, Kuruvila (2024) applied AHP integrated with RS and GIS data to generate flooding susceptibility outputs for Ernakulam District, India. However, traditional MCDA approaches commonly face challenges such as slow processing and difficulties in assigning accurate weights to influencing factors (Mudashiru et al., 2022), which may be compounded when integrated with ML algorithms (Shafik, 2025; Asiri et al., 2024; Liao et al., 2023).

FSM methodologies have gradually evolved from expert-driven frameworks to data-intensive approaches (Nurjahan et al., 2026). ML-based methods are capable of analyzing complex datasets and capturing intricate spatial and statistical relationships. ML algorithms successfully applied to FSM include RF (Megahed et al., 2023a; Tan et al., 2024; Rezvani et al., 2026), XGB (Abedi et al., 2022), GBM (Xu et al., 2023), Support Vector Machines (SVM) (Ali et al., 2024), Decision Trees (DT) (Band et al., 2020), Artificial Neural Networks (ANN) (Dahri et al., 2022), and Naïve Bayes (NB) (Abu El-Magd, 2022). These models exhibit strong performance in identifying flood-prone areas by leveraging advanced pattern recognition and data-processing capabilities in arid environments. Integrating MCDA with ML has proven effective for flooding assessment (Amiri et al., 2024; Adeyemi et al., 2024; Hussain et al., 2023; Bera et al., 2022; Megahed et al., 2024). Countries including Egypt, India, Pakistan, Saudi Arabia, and Greece have successfully implemented GIS-based MCDA frameworks to manage flash flood hazards (Shawky and Hassan, 2023; Subraelu et al., 2023; Abdullah et al., 2021; Roy et al., 2021). For example, Nachappa et al. (2020) evaluated the performance of two MCDA models against two ML algorithms for flood susceptibility prediction in Austria, concluding that ML models achieved superior results with AUC values exceeding 0.87.

In addition to hydrological extremes, arid coastal regions are highly vulnerable to desertification processes driven by vegetation decline, soil salinity, and unsustainable land use (Anwer et al., 2025; Bouzekri et al., 2023; Ghebrezgabher et al., 2019; Liu et al., 2018; Seboka, 2016; Yousef et al., 2020). These processes interact with flood hazards to amplify socio-economic exposure, underscoring the need for integrated multi-hazard frameworks. Multi-hazard susceptibility mapping has recently gained attention as a means of capturing compound risks, where overlapping hazards exacerbate vulnerability (Mudashiru et al., 2021; Pourghasemi et al., 2020; Edamo et al., 2022; Kim et al., 2025; Pathak, 2026).

Scenario-based analyses under Representative Concentration Pathways (RCPs) and Shared Socio-economic Pathways (SSPs) are increasingly applied to anticipate hazard evolution under climate and land-use change (van Vuuren et al., 2011; Völz and Hinkel, 2023; Fu et al., 2025; Feng et al., 2026; Liu et al., 2025; Xu et al., 2024; Wang et al., 2022). Aligning such projections with ecosystem restoration standards (Food and Agriculture Organization of the United Nations (FAO), 2025) ensures that scientific diagnostics translate into actionable resilience planning.

To overcome the limitations of conventional FSM techniques, this study proposes a hybrid methodology tailored to mountainous arid coastal regions. The approach combines AHP-based factor weighting with advanced ML algorithms and a dynamic flood inventory derived from Sentinel-1 SAR imagery. A distinctive feature of this framework is the generation of continuously updated flood inventories using Google Earth Engine (GEE) and Otsu thresholding for precise flood extent delineation (Otsu, 1979), enhancing the identification of inundated areas. Furthermore, SHAP analysis is applied to quantify the relative contribution of individual predictors, improving model interpretability a notable gap in FSM studies for mountainous arid coastal regions (Sahlaoui et al., 2021; Nohara et al., 2022; Yang et al., 2024). By employing SHAP, the study provides detailed insights into flood risk drivers in regions prone to recurrent flash floods due to distinctive geomorphology and extreme climatic conditions (Bashir and Alsalman, 2024), supporting targeted mitigation strategies and improved disaster management planning.

2 Area description

The study area spans a section of Egypt’s Red Sea coastline, extending from Wadi Al-Ambagi (34°17′05″E, 26°05′41″N), located south of El Quseir, to Wadi El-Barud (33°55′E, 26°45′39″N), situated approximately 2.6 km north of Safaga (Figure 1a). This region covers over 5,350 square kilometers and stretches roughly 88 km along the shoreline. The area is bounded to the east by the Red Sea and to the west by the watershed that separates the Red Sea drainage basin from the Nile and Mediterranean drainage systems.

Abdel Moneim, 2005), and (d) Main Wadis and Sub-basins with drainage channels.

Morphologically, the study area is bordered by W. Quseir El-Qudeim, Safaga, and Abu Shegeila in the north, W. El-Isewid in the south, and W. Nakheil and W. El Atshan in the west. The landscape presents a combination of rugged hills and flat coastal plains, with westward-tilting plateaus formed from Middle to Late Tertiary geological units. The region’s classification, which includes delineation of hydrographic basins, wadis, and alluvial fans (Figure 1b), is crucial for interpreting the topographic structure and informing hydrological and flood risk assessments. The interior is marked by highly uneven terrain, shaped by deformed Precambrian basement rocks and faulted blocks of Cretaceous to Eocene sediments, resulting in elevation variations of several hundred meters.

Extensive geological studies (Barron and Hume, 1902; Badawy, 2008; Bosworth et al., 2019) have documented the stratigraphy between Quseir and Safaga, identifying three principal rock units: basement rocks (including metavolcanics, metasediments, and granitoids), pre-rift sediments from the Mesozoic to Cenozoic eras, and syn-rift deposits spanning from the Late Oligocene to the present (Figure 1c). According to El Ramly (2010), El Gaby et al. (2017), and Hamimi and Abd El-Wahed (2020), the basement complex comprises diverse Pan-African lithologies such as gneisses, mafic intrusions, younger volcanic rocks, and granitic bodies. These Precambrian basement rocks are unconformably overlain by Phanerozoic strata, particularly in the southeastern portion of the study area. The pre-rift sequence begins with substantial deposits of Nubian Sandstone and continues with the Duwi, Dakhla, Quseir, and Esna formations (Mohamed, 2023; Moustafa and Khalil, 2019). The uppermost layer of this succession is represented by the Thebes Formation, composed predominantly of limestone and cherty limestone, dating from the Lower to Middle Eocene. Syn-rift deposits, which formed from the Late Oligocene to the Recent (Chiarella et al., 2024), start with coarse clastic units such as the Nakheil and Ranga formations and are subsequently overlain by evaporitic layers including the Um Mahra, Sayateen, and Abu Dabbab formations. Coastal exposures from the Late Miocene to the present reveal a stratigraphic sequence of carbonates, reef structures, clastic sediments, and evaporites (Alsharhan and Kendall, 2003).

From a hydrological perspective, the Red Sea Mountains are dissected by a network of deeply carved wadis that give rise to two distinct drainage systems. The Red Sea system is defined by short wadis and relatively small catchment areas, whereas the Nile-Mediterranean system comprises longer wadis and more expansive basins (Embabi, 2018). In the Red Sea system, the wadis form an external drainage network, flowing directly into the sea. Prominent examples, such as Wadis El Barud, Safaga, Queih, An-Nakheil, and Al-Ambagi, often traverse tectonic depressions filled with aeolian sands and alluvial deposits. The area’s complex geological framework and rugged terrain have shaped a dense and convergent drainage pattern, commonly exhibiting dendritic and trellised configurations. These systems include both consequent and obsequent wadis, reflecting the structural and topographic controls on water flow (Figure 1d).

3 Methodology

Figure 2 presents the flowchart outlining the complete workflow adopted for FSM in the Safaga–El Quseir region. The approach involves several critical stages, beginning with the generation of a flood inventory map and the selection of flooding-effecting variables. These factors were assessed using statistical techniques such as Pearson correlation and ordinary least squares (OLS) regression to determine their relevance. The modeling phase incorporated multiple ML algorithms (FR, XGB, and GBM), each integrated with Analytical Hierarchy Process (AHP) weightings to enhance predictive accuracy. Model performance was evaluated using a suite of validation metrics, including the kappa coefficient, sensitivity, AUC value, positive predictive value (PPV), and negative predictive value (NPV). To improve interpretability, SHAP analysis was applied, offering insights into the individual contribution of each input factor to the model’s predictions. To evaluate model robustness and reduce potential selection bias associated with a single holdout split, a 5-fold cross-validation (CV) procedure was additionally performed for the RF, XGB, and GBM models. The dataset was randomly partitioned into five equal subsets, where four folds were used for training and one-fold for validation in each iteration. Model performance was assessed using Accuracy and AUC metrics, and the final results were reported as the mean and standard deviation across the five folds.

3.1 Flood inventory map generation

In this study, flood inventory maps for the Safaga–El Quseir region were generated by combining GPS-based field observations with Sentinel-1 satellite imagery. The analysis utilized pre-processed Sentinel-1 C-band (VV/VH) data, accessed via GEE platform, covering the period from February 2016 to October 2025 (https://www.youm7.com). To enhance image quality, speckle noise was minimized through filtering techniques. Flood events were identified by comparing pre- and post-event backscatter values, applying Otsu’s automatic thresholding algorithm. This method assumes a bimodal histogram and determines an optimal threshold to classify pixels as flood (1) or non-flood (0). The resulting flood inventory was cross-validated with historical flood records and refined to include 200 representative locations, including 100 verified flood locations and 100 non-flood control points to ensure balanced class representation. The flood inventory dataset derived from Sentinel-1 SAR imagery and historical flood records (2016–2025) was randomly divided into training (70%) (140 points) and testing (30%) (60 points) subsets. Therefore, the validation conducted in this study represents an internal validation framework rather than a fully independent temporal or spatial validation. A detailed time series of flood occurrence maps spanning 2016 to 2025 is presented in Figure 3.

3.2 Identification of flood susceptibility factors

To facilitate the construction of FSM models for the Safaga–Quseir region, a set of flood-influencing variables was identified based on well-established studies (e.g., Islam et al., 2025; Shah and Ai, 2024; Latif and He, 2024; Bashirgonbad et al., 2024; Mudashiru et al., 2021). These factors were selected for their proven relevance in representing the hydrological, topographic, geological, and land-use characteristics that contribute to flood events. To ensure their suitability and statistical robustness, the variables were assessed using OLS regression, multicollinearity diagnostics, and the Information Gain Ratio (InGR) algorithm. These analytical techniques helped eliminate redundant inputs and retain only the most influential predictors for model development. Table 1 provides a detailed overview of the selected flood-conditioning parameters, including their data sources and descriptive statistics.

3.3 Flooding susceptibility factors

3.3.1 Topographic factors

Topographic variables extracted from the Digital Elevation Model at 30-m resolution using ArcGIS played a pivotal role in evaluating flood susceptibility across the Safaga–El Quseir basins (Rawat et al., 2012; Munir et al., 2019; Leal et al., 2022; Taherizadeh et al., 2023). These factors were processed through the GEE platform to ensure spatial consistency and analytical precision. Elevation within the study area ranges from sea level to approximately 1,400 m (Figure 4a). Lower elevations, particularly those below 190 m, were found to be more vulnerable to flooding, while areas above 640 m exhibited reduced susceptibility. Slope values span from 0° to 55° (Figure 4b), with gentle slopes in the basin’s lower regions contributing to increased flood risk due to slower runoff and greater water retention potential. Aspect was categorized into ten directional classes (Figure 4c), influencing microclimatic conditions and soil moisture dynamics, which in turn affect infiltration and surface runoff. TRI values ranged from 0.54 to 0.3 (Figure 4d), with flatter zones, especially those near wadis and depressions, showing higher flood susceptibility due to lower ruggedness. Conversely, more rugged terrain with elevated TRI values was less prone to flooding. Distance to roads varied from 0 to 43,000 m (Figure 4e). Areas within 0–8,700 m of road infrastructure were more exposed to flood hazards, largely due to concentrated overland flow, reduced infiltration, and the disruption of natural drainage pathways by road embankments. In contrast, regions farther from roads experienced lower direct flood impacts, as runoff was more evenly distributed across broader catchments. These findings underscore the importance of integrating road network data into flood susceptibility modeling to better capture anthropogenic influences on hydrological behavior.

3.3.2 Hydrologic factors

Hydrological factors were crucial in assessing flood susceptibility across the Safaga–Quseir region. TWI derived from flow accumulation using SRTM DEM with calculations based on Equation 1, ranged from 21 to 28. Higher TWI values (>21) indicated areas with greater potential for water accumulation and, therefore, increased flood risk (Figure 4f). Drainage density calculated using GIS line density tool and varied between 0.14 and 5 km/km² (Figure 4g). Regions with high drainage density (0.4–5 km/km²), typically in middle and lower catchments, showed elevated flood risk due to intensified surface runoff. In contrast, areas with sparse drainage networks were less susceptible. Runoff played a primary role in flood triggering, with annual precipitation (2016–2025) obtained from TRMM data and interpolated using the IDW tool. Volume runoff values ranged from 250,000 to 16,000,000 m³, with higher flood susceptibility observed in lower basin zones where runoff exceeded 16,000,000 m³ (Figure 4h). Additionally, distance from streams, computed using Euclidean distance based on ArcHydro, ranged from 0 to 26,000 m (Figure 4i). Zones located within 5,100 m from streams showed a significantly higher flood hazard, while areas farther away experienced reduced susceptibility. Together, these factors highlighted the spatial variability of flood risk and the importance of integrating topographical and hydrological factors for accurate flood modeling in arid coastal basins such as the Safaga- EL-Quseir region.

As denotes the catchment area, and β denotes the slope. Locations with higher TWI values are more prone to flooding compared to areas with lower values.

3.3.3 Terrain factors

Soil types, sourced from the SoilGrids datasets (https://soilgrids.org) and IUSS Working Group (2015), included basement, clastic, non-clastic, and wadi deposits (Figure 4j). The clastic soil in the lower part exhibited high runoff potential and reduced infiltration, leading to increased flood susceptibility. STI quantifies the potential of surface flow to transport sediments, reflecting both slope steepness and upstream contributing area. STI layer was generated using Equation 2 from Moore and Burch (1986) and processed with QGIS software (Figure 4k).where (As) is the specific catchment area and (β) is the slope angle.

Higher STI values (<3400) (Figure 4k) indicate zones with greater runoff power and sediment transport capacity, which often correspond to flood-prone channels and lower basin areas. Therefore, STI plays a key role in identifying areas with high flood susceptibility and erosion potential.

3.4 Information gain ratio analysis

Before training and validating the flood susceptibility models, it was necessary to assess the relative importance of the flood-conditioning parameters. This was achieved using InGR analysis, implemented in Python via the scikit-learn library within a Google Collaboratory environment, calculated by Equations 3–7 (Khodaei et al., 2025). The InGR method serves as a robust metric for feature selection, quantifying the contribution of each variable to reducing uncertainty in the dataset (Mueller et al., 2023). It is computed by dividing the Information Gain (IG), which measures the reduction in entropy after a split, by the Split Information (SI), which represents the intrinsic information of a given feature. Each parameter was assigned an InGR score, with higher values indicating a stronger influence on flooding (Talukdar et al., 2020; Arora et al., 2021; Nguyen et al., 2023; Islam et al., 2023). The application of InGR in this study enabled the identification of the most predictive parameters, supporting more efficient and accurate model development. In addition, entropy-based metrics were utilized during decision tree construction to determine optimal data splits, with entropy values (0-1), providing further insight into data structure and classification performance.where IG (X, F) represents the information gain of factor F, Entropy(X) is the entropy before splitting, Entropy(X/F) is the conditional entropy after splitting, X_j ​ denotes the subset generated by factor F, m represents the total number of subsets and p_i represents the probability of occurrence of class i (flood and non-flood classes) in the dataset. The Information Gain Ratio (InGR) evaluates the relative importance of flood-conditioning factors by normalizing the information gain using split entropy.

3.5 AHP weightage estimation

AHP is a widely used method for MCDA because it effectively addresses complex decision-making problems (Saaty, 1980). It has been extensively applied in the assessment and prediction of natural hazards, including FSM (e.g., Chakraborty and Mukhopadhyay, 2019; Lin et al., 2020; Spanidis et al., 2021; Abdrabo et al., 2023; El-Rawy et al., 2023; Zhran et al., 2024). In flood susceptibility studies, AHP proves invaluable in determining the relative importance of contributing factors by integrating expert opinions and GIS-based environmental data to support informed decision-making (Uruvila, 2024).

Eleven factors influencing floods were initially considered to identify flood-prone areas based on expert judgment and previous literature: elevation, slope, aspect, TRI, distance to roads, TWI, drainage density, volume runoff, distance to streams, soil types, and STI. Pairwise comparison matrices were developed for each factor (Figure 5; Table 2), using a scale from 1 to 5 to reflect the relative importance of each criterion, where 1 indicates low significance and 5 represents high importance of one factor over another. Reciprocal values (e.g., 1/3, 1/5) were also used to express lesser importance.

Following the approach outlined by Penki et al. (2023), the pairwise comparison matrix was normalized by dividing each element by the column sum, and the criteria weights were then calculated as the arithmetic mean of each row. These weights were subsequently used to compute the percentage contribution of each thematic layer. The final flood susceptibility map was produced by multiplying each standardized criterion map by its corresponding weight and summing the results across all layers (Mahmoud and Gan, 2018). The consistency of judgments in the pairwise comparisons was evaluated using the Consistency Ratio (CR) (Equation 8), as described in previous studies (Nguyen and Hoang, 2023), ensuring the reliability of the MCDA process.

CI is the Consistency Index, and RCI refers to the Random Consistency Index.

3.6 Machine learning models application

Flood inventory data derived from Sentinel-1 imagery were used as the dependent variable, with binary classification assigning a value of 1 to flood-affected locations and 0 to non-flooded areas. The dataset was refined to include 200 representative points, which were split into two subsets: 70% (140 points) for training and 30% (60 points) for testing. Similar sample sizes have been widely applied in flood susceptibility studies conducted in data-scarce arid and semi-arid regions. Although the dataset size is relatively limited, balanced class representation was maintained to reduce classification bias, and ensemble ML models demonstrated stable predictive performance under data-scarce conditions. A comprehensive set of independent variables was incorporated into the modeling process, including elevation, slope, aspect, TRI, distance to roads, TWI, drainage density, runoff, distance to streams, soil types, and STI. These variables capture the hydrological, topographical, and geological conditions influencing flood occurrence. To evaluate flood susceptibility, a series of hybrid ensemble ML models were developed: RF, XGB, and GBM models each integrated with AHP weightings, resulting in AHP-RF, AHP-XGB, and AHP-GBM models. Model hyperparameters were optimized using a grid search approach combined with k-fold cross-validation within the ‘caret’ package in R (version 3.4). A range of candidate values was tested for each model parameter, and the optimal configuration was selected based on model performance, particularly the Area Under the Receiver Operating Characteristic Curve (AUC-ROC). This approach ensures robust model calibration, reduces overfitting, and enables a fair comparison among the applied models.

3.6.1 Random forest (RF)

Random Forest (RF) is an ensemble machine learning algorithm based on the construction of multiple decision trees using bootstrap aggregation (bagging) (Breiman, 2001). Each decision tree is trained on a random subset of the training data and a random subset of predictor variables. The final classification is obtained by aggregating the predictions of all individual trees through a majority voting process. RF is widely used in classification problems due to its high accuracy, robustness to overfitting, and ability to handle nonlinear relationships between input variables and target classes (Waqas et al., 2021; Tariq et al., 2022; Majeed et al., 2023).

3.6.2 Extreme gradient boosting (XGB)

XGB model is a widely used tree-based ML algorithm known for its high performance and efficiency (Noorunnahar et al., 2023). It is particularly valued for its scalability, robustness in handling various data types and nonlinear patterns, built-in feature selection, and model interpretability. It constructs predictive models by sequentially combining multiple weak decision trees and applying gradient descent to minimize the loss function (Shehadeh et al., 2021; Brunner et al., 2021). This results in faster processing and broad applicability across both classification and regression problems.

3.6.3 Gradient boosting machine (GBM)

The GBM model was applied to supervised ML tasks to address both classification and regression problems (Mirzaei et al., 2021). GBM builds an ensemble model by sequentially combining simple base learners to improve predictive performance (Sibindi et al., 2023). The method centers on three key components: optimizing a designated loss function and iteratively refining those predictions using an additive process to reduce the loss. The GBM model was tuned using the Bernoulli loss function across 100 boosting iterations. Of the eleven input predictors, all were found to have a non-zero contribution to the FSM model.

3.7 Hyperparameter settings and models training workflow

To ensure reproducibility and transparency, all machine learning models (Random Forest, XGBoost, and Gradient Boosting Machine) were implemented using Python-based libraries, including scikit-learn and XGBoost. All experiments were conducted using a fixed random seed (random_state = 42) to ensure consistent and reproducible results across all runs. The hyperparameters for each model were defined as follows: Random Forest (RF) was configured with 200 trees (n_estimators = 200), unlimited tree depth (max_depth = None), a minimum samples split of 2, a minimum samples leaf of 1, and the Gini impurity criterion. For the XGBoost (XGB) model, the parameters included n_estimators = 200, learning_rate = 0.1, max_depth = 6, subsample = 0.8, colsample_bytree = 0.8, and the binary classification objective function. The Gradient Boosting Machine (GBM) model was set with n_estimators = 200, learning_rate = 0.1, max_depth = 3, subsample = 0.8, and deviance as the loss function. No extensive hyperparameter optimization techniques such as grid search were applied; instead, the parameters were selected based on commonly adopted values in previous flood susceptibility studies and refined through iterative testing to achieve stable model performance. The overall modeling workflow consisted of data preprocessing and normalization, derivation of AHP weights, extraction of training and testing samples using a 70/30 split, model training, and performance evaluation using multiple statistical metrics including Accuracy, AUC, Cohen’s Kappa, PPV, and NPV.

3.8 AHP-ML approach

This study introduced a novel hybrid methodology that integrates the AHP method with various ML models for flooding susceptibility. Research has shown that combining multiple approaches is more effective than relying on a single ML algorithm for FSM analysis (Fernández et al., 2024). The integration of AHP with machine learning improves flood susceptibility modeling (FSM) by combining expert-based factor weighting with data-driven predictive algorithms (Schmidt et al., 2020; Hasan et al., 2025; Mishra and Prasad, 2024). In this framework, AHP was first applied to derive weights for each flood-conditioning factor based on pairwise comparisons. The thematic layers were then normalized and multiplied by their corresponding AHP weights to generate weighted raster layers. These weighted layers were subsequently used as predictor variables, from which training samples (flood and non-flood points) were extracted and used as input for the machine learning models (RF, XGB, and GBM). This procedure ensures that the ML models learn from spatial patterns that already incorporate expert knowledge of factor importance. This hybrid approach enhances model performance by integrating the interpretability of AHP with the predictive capability of machine learning algorithms, making it particularly suitable for data-scarce environments.

3.9 Validation analysis

The predictive performance of the models was assessed using reliable and practical experimental methods and essential evaluation equations (accuracy, the kappa coefficient, AUC, PPV, and NPV), similar to previous studies (Liu et al., 2023; Duan et al., 2022; Pham et al., 2020; Costache et al., 2019; Pham et al., 2020; Al-Juaidi, 2023), which were used as the primary indicators of model effectiveness (Equations 9–12).

Accuracy measures the overall proportion of correctly classified samples and was calculated as:

The Cohen’s Kappa coefficient evaluates the agreement between observed and predicted classifications while accounting for agreement occurring by chance:where Pobs is the observed agreement and Pexp is the expected agreement due to random chance.

The Positive Predictive Value (PPV) represents the proportion of correctly predicted flood locations among all predicted positive samples:

The Negative Predictive Value (NPV) indicates the proportion of correctly predicted non-flood locations among all predicted negative samples:where TP represents true positives, TN represents true negatives, FP represents false positives, and FN represents false negatives.

The 5-fold cross-validation results confirmed the stability and robustness of the hybrid ML models. Among the evaluated approaches, the RF model achieved the highest average predictive performance (Accuracy = 0.964 ± 0.005; AUC = 0.95 ± 0.01) (Table 3), followed by XGB and GBM. The relatively small standard deviation values indicate consistent model behavior across different validation folds and suggest limited sensitivity to dataset partitioning.

3.10 SHAP analysis

SHAP was used to interpret the contribution of individual features to the predictions developed by the ML models (Aydin and Iban, 2023; Wang et al., 2023; He et al., 2025). SHAP values were used to measure each feature’s impact on predictions. The top-performing RF model was analyzed using Python’s SHAP library on Google Colab. SHAP applies an additive framework from cooperative game theory (Lundberg and Lee, 2017), producing Shapley values by aggregating feature contributions (Equation 13).where, z′∈ {0,1} M denotes the coalition of contributing features, M the total number of features, and ϕj ∈R represents the Shapley value for a variable j.

In this study, SHAP interpretation was conducted through multiple visual tools (waterfall plots, decision plots, heat maps, and force plots). SHAP values are centered around the model, showing the strength of each feature’s effect. A positive SHAP value indicates that a feature increases the likelihood of higher flood susceptibility, and a negative value suggests it reduces the possibility of lower flood susceptibility. Higher values reflect greater influence on the model output.

3.11 Development of the environmental sensitivity and desertification index (ESDI)

To complement flood susceptibility analysis, an Environmental Sensitivity and Desertification Index (ESDI) was developed by integrating vegetation dynamics, soil degradation proxies, aridity indicators, and land-use intensity. Vegetation indices such as NDVI and SAVI were derived from multi-temporal satellite imagery, which have proven effective in monitoring land degradation and desertification processes in arid and semi-arid regions (Kocur-Bera and Małek, 2024; Seboka, 2016; Liu et al., 2018). Soil degradation layers were constructed using Landsat-based assessments of desertification (Ghebrezgabher et al., 2019) and regional sandy desertification mapping (Bouzekri et al., 2023). Land-use intensity was quantified using settlement expansion and agricultural activity proxies, while aridity indices were calculated from precipitation and temperature records. Each thematic layer was standardized and subjected to Analytic Hierarchy Process (AHP) weighting to derive a composite raster representing desertification sensitivity, following established multi-criteria frameworks (Anwer et al., 2025; Yousef et al., 2020).

3.12 Multi-hazard susceptibility framework

The study advances beyond single-hazard flood modeling by integrating ESDI with flood susceptibility maps into a multi-hazard framework. Weighted thematic layers were combined within hybrid AHP–machine learning models (RF, XGB, GBM), enabling simultaneous evaluation of flood and desertification susceptibility. This approach builds upon recent advances in multi-hazard modeling that employ machine learning to capture complex interactions between environmental drivers (Mudashiru et al., 2021; Edamo et al., 2022; Kim et al., 2025). Predictor importance was quantified using Shapley Additive Explanations (SHAP), allowing transparent comparison of flood-related versus desertification-related drivers. Similar SHAP-based interpretability has been successfully applied in multi-hazard contexts (Yousef et al., 2020; Pourghasemi et al., 2020; Pathak, 2026), reinforcing the methodological novelty of this integrated framework.

3.13 Scenario-based susceptibility analysis

To evaluate future system dynamics, scenario-based analyses were conducted. Climate projections under RCP 4.5 and RCP 8.5 were incorporated to simulate runoff and temperature variability, following the Representative Concentration Pathways framework (van Vuuren et al., 2011). Recent advances in climate adaptation scenario classification informed the integration of multiple pathways into hazard modeling (Völz and Hinkel, 2023). Land-use change scenarios reflecting urban expansion and tourism growth were modeled using Shared Socioeconomic Pathways (SSPs), which have been widely applied in urban flood susceptibility and resilience studies (Fu et al., 2025; Feng et al., 2026; Liu et al., 2025; Xu et al., 2024). The resulting susceptibility maps illustrate the evolution of hazard exposure under different environmental stressors, thereby enhancing the predictive and adaptive capacity of the framework.

3.14 Validation and comparative assessment

Validation of desertification outputs was performed against FAO’s State of the World’s Land and Water Resources for Food and Agriculture (Food and Agriculture Organization of the United Nations (FAO), 2025) and UNCCD’s World Atlas of Desertification. Comparative analyses were undertaken by applying the integrated framework to an additional arid region to demonstrate transferability, consistent with prior multi-hazard validation approaches (Ghebrezgabher et al., 2019). Model robustness was further evaluated using k-fold cross-validation, spatial independence checks, and reproducibility metrics, ensuring methodological rigor and generalizability.

3.15 Socio-economic and policy integration

To bridge scientific outputs with practical decision-making, susceptibility maps were overlaid with socio-economic datasets including population density, infrastructure distribution, and tourism assets. This integration highlights zones of heightened exposure and informs targeted interventions. Policy-oriented recommendations were derived for urban planning, ecosystem restoration, and disaster risk reduction, building upon recent studies that emphasize socio-economic overlays in hazard resilience (Fu et al., 2025; Feng et al., 2026; Wang et al., 2022). By explicitly linking hazard susceptibility with human and economic systems, the framework is positioned as a decision-support tool for multi-hazard resilience in arid regions.

3.16 Validation of ESDI and multi-hazard maps

Validation of the Environmental Sensitivity and Desertification Index (ESDI) was performed by comparing the classified sensitivity zones with reference datasets from the FAO Global Land Degradation Information System and the UNCCD Desertification Atlas. A confusion matrix was constructed to evaluate agreement between the ESDI map and the reference data (Filianoti et al., 2020). Accuracy indices including overall accuracy, kappa coefficient, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated.

For the multi-hazard susceptibility map, validation was conducted by overlaying compound susceptibility zones with independent hotspot datasets (flood records and desertification risk maps). Agreement was quantified using kappa statistics and ROC–AUC values to assess the reliability of the integrated framework.

4 Results

4.1 Multicollinearity and OLS analysis

Multicollinearity and OLS analyses were conducted for all eleven variables (Table 4). The variance inflation factor (VIF) for each factor was below 10, suggesting the absence of severe multicollinearity and confirming their suitability for model inclusion. Among them, elevation recorded the highest VIF (5.807), while TRI had the lowest (1.03). OLS results showed that aspect, distance to roads, TWI, drainage density, runoff, and STI were highly significant predictors of flood susceptibility mapping (FSM) (p < 0.001). Elevation, slope, and soil type also showed statistically significant relationships with FSM (p < 0.01).

4.2 Pearson’s correlation and information gain ratio analysis

Pearson’s correlation analysis between flood occurrence and its potential drivers (Figure 6a) revealed that stronger correlations are linked to a higher flood probability, consistent with previous findings (Douglas et al., 2000). Key influencing factors include elevation, TRI, slope, TWI, and soil types. In the study area, flooding primarily occurred in regions with clayey soils, elevations below 350 m, low TRI values, slopes of less than 5°, TWI values ranging from 9.3 to 14, and flat to gently sloping terrains. Figure 6b displays the calculated InGR values for each parameter. The results show elevation (0.480), volume runoff (0.355), slope (0.341), TWI (0.253), and aspect (0.259) as the most influential factors, with the highest InGR values among all factors. Moderate importance factors include TRI (0.084), distance to streams (0.060), and soil types (0.053). Conversely, drainage density (0.034), STI (0.025), and distance to roads (0.015) have minimal impact on FSM prediction.

4.3 Flood susceptibility maps

An AHP Pairwise Comparison Matrix (PCM) model was applied to derive normalized weights for each criterion. The AHP analysis ranked elevation as the most influential factor (0.198), followed by runoff, slope, drainage density, soil types, TWI, TRI, aspect, distance from streams, and distance to roads in descending order of importance. To predict FSM, hybrid AHP–ML models were implemented in ArcGIS 10.8. Each model generated FSM values for every pixel across the basins (Figures 7a–d). The results were categorized into four susceptibility classes using the natural breaks classification method in ArcGIS.

For the AHP model, 42.94% of the area was divided into the very low susceptibility class, low (29.8%), moderate (26.56%), and high (0.7%) (Figure 7a). For the AHP-RF model, 43% of the area was categorized into the very low susceptibility class, low (25%), moderate (30.7%), and high (1.3%) (Figure 7b). In the AHP-XGB model, there were 42.85% (very low), 27% (low), 29.45% (moderate), and 0.7% (high) (Figure 7c). For the AHP-GBM model, areas with very low susceptibility accounted for 43.09%, while those with low susceptibility covered 35.1% of the area, 21.7% (moderate), and 0.11% (high) (Figure 7d). All susceptibility maps, zones with flat terrain and substantial runoff, particularly in the lower parts of the studied region, were consistently classified as moderate to high flood risk areas.

4.4 Validation analysis results

The AHP-RF model achieved the highest testing accuracy (0.95), surpassing AHP-XGB (0.925), AHP-GBM (0.90), and AHP (0.85). Figure 8a compares the AUC values for all models used for FSM prediction. All three of the AHP-ML models demonstrated strong predictive capability; the AHP-RF model achieved the highest AUC (0.95), making it the most effective in this study area. Kappa scores a round from 0.881 to 0.936, further supporting model robustness. The AHP-XGB model displayed the highest sensitivity, whereas AHP-GBM achieved the highest specificity. Additionally, both AHP-GBM and AHP-XGB demonstrated superior predictive value, with PPV and NPV exceeding 0.94 (Figure 8b).

4.5 SHAP analysis results

The study used an interpretable framework to clarify model performance and variable influence. using the summary plot revealed that elevation, volume runoff, and drainage density were the most influential predictors across all flood susceptibility models. Variables such as soil type, aspect, stream proximity, and TWI showed moderate influence, while slope, TRI, and road distance were among the least impactful (Figure 9a). Insights from the waterfall plot indicated that volume runoff, elevation, stream distance, and the Stream Power Index (STI) positively contributed to flooding susceptibility. Conversely, drainage density, aspect, soil type, and TWI were associated with reduced susceptibility (Figure 9b). The SHAP heatmap emphasized volume runoff, elevation, and TWI as key drivers of model accuracy (Figure 9c). Pixel‐wise accuracy analysis (Figure 9d) revealed that volume runoff, elevation, TWI, and drainage density most frequently improved classification performance for flood susceptibility mapping. The force plot (Figure 9e) illustrates the decision‐making process, showing that volume runoff, elevation, TRI, STI, and distance from the streams all enhance classification accuracy, while TWI, drainage density, and aspect reduce it.

4.6 Environmental sensitivity and desertification index (ESDI)

To complement the flood susceptibility analysis, an Environmental Sensitivity and Desertification Index (ESDI) was developed by integrating vegetation proxies (NDVI, SAVI), soil degradation indicators, aridity measures, and land-use intensity. Vegetation indices have been widely applied to monitor land degradation in arid and semi-arid regions (Kocur Bera and Małek, 2024; Seboka, 2016; Liu et al., 2018). Soil degradation layers were informed by Landsat-based desertification assessments in Eritrea and Algeria (Ghebrezgabher et al., 2019; Bouzekri et al., 2023). Land-use intensity was quantified using settlement expansion and agricultural activity proxies, while aridity indices were derived from precipitation and temperature records. The composite ESDI map (Table 5; Figure 10) was classified into five sensitivity zones very low, low, moderate, high, and very high revealing that 38% of the area falls under very low sensitivity, 27% under low, 25% under moderate, 8% under high, and 2% under very high. The most sensitive zones occur in coastal sabkhas, saline depressions, and intensively cultivated floodplains, where vegetation loss and soil degradation coincide with high aridity. Conversely, elevated plateaus and reclaimed lands exhibit low sensitivity due to stable vegetation and resilient soils. Spatial hotspots of desertification risk align with rapid urban expansion and tourism development along the Safaga–El Quseir corridor, consistent with regional studies identifying vegetation decline and soil degradation as key drivers of desertification in North African arid landscapes (Anwer et al., 2025; Yousef et al., 2020).

4.7 Multi-hazard susceptibility mapping

To advance beyond single-hazard analysis, the flood susceptibility maps were integrated with the Environmental Sensitivity and Desertification Index (ESDI) to construct a multi-hazard susceptibility framework. The overlay analysis revealed distinct spatial patterns of compound risk across the Safaga–El Quseir region. The integrated map (Table 5; Figure 11) classified the area into five compound susceptibility zones very low, low, moderate, high, and very high indicating that 32% of the region falls under very low susceptibility, 28% under low, 27% under moderate, 10% under high, and 3% under very high. The highest compound susceptibility occurs along coastal floodplains, sabkhas, and intensively cultivated lands, where flood risk overlaps with severe desertification sensitivity, while elevated plateaus and geomorphically stable terrain exhibit minimal susceptibility. Rapid land-use change along the urban and tourism corridors of Safaga and El Quseir intensifies both flood exposure and desertification vulnerability, consistent with recent multi-hazard studies integrating hydrological and land-degradation drivers within machine-learning frameworks (Mudashiru et al., 2021; Edamo et al., 2022; Kim et al., 2025; Pourghasemi et al., 2020). SHAP analysis confirmed elevation, runoff volume, drainage density, NDVI, and aridity index as the most influential predictors of compound susceptibility, demonstrating that hydrological and ecological drivers jointly determine risk in arid coastal systems consistent with multi-hazard modeling approaches in other mountainous and semi-arid regions (Yousef et al., 2020; Pathak, 2026).

4.8 Scenario-based susceptibility analysis

To evaluate future dynamics of hazard exposure, scenario-based analyses were conducted using climate projections under Representative Concentration Pathways (RCP 4.5 and RCP 8.5) (Figure 12) and socio-economic trajectories defined by Shared Socioeconomic Pathways (SSPs) (Table 6). These scenarios were integrated into the multi-hazard framework to simulate changes in flood susceptibility and desertification sensitivity under varying climate and land-use conditions.

4.8.1 Climate forcing (RCP scenarios)

Under RCP 4.5, moderate increases in runoff variability and temperature resulted in a 7% expansion of high-risk zones, primarily in lower catchments and coastal floodplains. In contrast, RCP 8.5 produced a more pronounced shift, with 15% of the study area transitioning into high or very high susceptibility classes, reflecting intensified aridity and runoff concentration. These findings are consistent with global projections of hazard amplification under higher emission pathways (van Vuuren et al., 2011; Völz and Hinkel, 2023).

4.8.2 Land-use change (SSP scenarios)

Socio-economic simulations revealed that SSP2 (middle-of-the-road development) maintained moderate hazard expansion, while SSP3 (regional rivalry with weak environmental governance) exacerbated compound risk, particularly along the Safaga–El Quseir tourism corridor. Under SSP5 (fossil-fuel intensive growth), rapid urban expansion and infrastructure development led to a 12% increase in high-risk zones, highlighting the vulnerability of reclaimed lands and sabkhas to both flooding and desertification. These results align with recent scenario-based urban flood studies (Fu et al., 2025; Feng et al., 2026; Liu et al., 2025; Xu et al., 2024).

4.8.3 Integration with FAO ecosystem restoration standards

The scenario outputs were cross-referenced with the FAO Standards of Practice to Guide Ecosystem Restoration (Food and Agriculture Organization of the United Nations (FAO), 2025), currently under review. This comparison emphasized that restoration priorities such as vegetation recovery, soil stabilization, and sustainable land management are most urgently needed in zones projected to shift into high or very high susceptibility under RCP 8.5 and SSP3/SSP5. The integration of FAO benchmarks ensures that scenario-based hazard mapping is not only predictive but also aligned with international restoration and resilience frameworks.

4.8.4 Validation of ESDI and multi-hazard maps

The ESDI map validation against FAO/UNCCD datasets yielded an overall accuracy of 87% and a kappa coefficient of 0.71, indicating substantial agreement. Sensitivity and specificity values were 0.84 and 0.89, respectively. The confusion matrix is presented in Table 7.

From this matrix, PPV was calculated as 0.88 and NPV as 0.87, confirming balanced predictive performance across sensitivity classes.

For the multi-hazard susceptibility map, overlay analysis with independent hotspot datasets produced a kappa coefficient of 0.68 and ROC–AUC of 0.90, confirming that the integrated framework reliably identifies compound risk zones.

5 Discussion

5.1 Comparison of AHP–ML model performance with previous studies

Modeling and simulation platforms are critical approaches for flood hazard assessment, enhancing decision-makers’ abilities and supporting the creation of effective environmental policies (Efraimidou and Spiliotis, 2024; Levy, 2025). The diversity in their designs and features produces varying levels of predictive accuracy, making a wide range of modeling options valuable for policy development. In different regions, diverse methods have been independently applied for flood susceptibility mapping (FSM).

For instance, Ma et al. (2021) applied the XGB model in Yunnan Province, China, whereas Majeed et al. (2023) used the RF model to map flood vulnerability in the Jhelum region of Punjab Province, Pakistan. Likewise, Tariq et al. (2022) implemented the AHP model in the Chaj Doab region between the Jhelum and Chenab Rivers. These variations emphasize the importance of comparative studies that evaluate different approaches under consistent environmental conditions to objectively assess model performance (Siau and Rossi, 2011).

The present study addresses this gap by comparing the predictive capabilities of hybrid AHP–ML approaches and traditional MCDA techniques in the Safaga–El Quseir region of Egypt. The evaluated models included AHP-RF, AHP-XGB, and AHP-GBM. Among these approaches, the AHP-RF model achieved the highest predictive performance with an AUC value of 0.95. This result exceeds the predictive accuracies reported for traditional AHP-based FSM studies, which generally range between 0.70 and 0.85 (Salma et al., 2023; Vujović et al., 2025).

The superior performance of the hybrid AHP–ML models can be attributed to their ability to integrate expert-based weighting with data-driven learning capabilities. ML algorithms reduce the uncertainty associated with subjective weighting and can better capture nonlinear relationships between flood-conditioning factors and flood occurrence (Herrera et al., 2022).

Several previous studies have also demonstrated the effectiveness of ensemble and hybrid approaches for FSM. Costache et al. (2021) showed that ensemble statistical models outperform standalone techniques. Similarly, Riaz et al. (2024) introduced SVM-based ensemble frameworks combined with weights-of-evidence and reported improved predictive performance. Akiner and Ghasri (2024) further highlighted the benefits of integrating adaptive neuro-fuzzy inference systems with nature-inspired optimization algorithms. Consistent with these findings, the current study confirms that coupling AHP with ML algorithms significantly improves FSM prediction accuracy in arid environments.

5.2 Key flood conditioning factors and their influence

The results of the InGR and SHAP analyses revealed that hydro-topographic parameters play dominant roles in shaping flood susceptibility within the study area. InGR analysis identified elevation (0.480), volume runoff (0.355), slope (0.341), TWI (0.253), and aspect (0.259) as the most influential flood-conditioning factors. These findings are consistent with previous studies that recognized elevation, runoff characteristics, and terrain morphology as primary determinants of flood occurrence in arid and semi-arid environments (Megahed et al., 2023b; c; Nkeki et al., 2022). The SHAP analysis provided additional insight into the contribution of each variable to model predictions. Elevation, volume runoff, and drainage density showed the strongest positive influence on flood susceptibility, highlighting the importance of terrain configuration and hydrological connectivity in controlling runoff concentration and flood propagation. Areas with low elevation and high runoff accumulation are naturally more prone to water convergence and inundation.

Moderate contributions were observed for soil type, aspect, distance to streams, and TWI, indicating that these factors influence local infiltration, flow direction, and water retention capacity. In contrast, slope, TRI, and distance to roads exhibited relatively lower impacts on model performance. Furthermore, SHAP results demonstrated that volume runoff, elevation, and STI positively contributed to flood susceptibility, whereas drainage density, TWI, and aspect reduced susceptibility in certain areas. These findings collectively emphasize that flood generation in the Safaga–El Quseir region is primarily controlled by the interaction between terrain morphology, runoff accumulation, and hydrological characteristics.

5.3 Spatial distribution of flood susceptibility zones

The generated FSM maps revealed a strong spatial agreement among the hybrid AHP–ML models regarding the distribution of flood-prone zones. Areas characterized by low elevation, gentle slopes, elevated TWI values, and proximity to drainage networks consistently exhibited moderate to high flood susceptibility, which agrees with the observations of Islam et al. (2023).

Among the tested approaches, the AHP-RF model produced the most reliable spatial prediction results. Approximately 1.3% of the study area was classified as highly susceptible to flooding, whereas 30.7% was categorized as moderately susceptible. These susceptible zones are mainly concentrated in the middle and lower basin sections, particularly along the northwest–southeast direction of the study area.

Validation against historical flood hazard layers from 2016 to 2025 further confirmed the reliability of the predicted susceptibility zones. Wadi Hamraween and Wadi Ambagi were consistently identified as high-susceptibility areas, reflecting their geomorphological and hydrological characteristics. Moderate to high susceptibility was also observed in Wadi Naqarah and Wadi Abu Shiqily. Conversely, Wadi Barud and Wadi Safaga predominantly showed low to moderate susceptibility levels. The flood inventory dataset derived from Sentinel-1 SAR imagery and historical flood records (2016–2025) was randomly divided into training (70%) and testing (30%) subsets. Therefore, the validation conducted in this study represents an internal validation framework rather than a fully independent temporal or spatial validation.

The observed spatial variability in flood susceptibility provides valuable information for regional planning and hazard mitigation. Identifying highly vulnerable basins can assist decision-makers in prioritizing infrastructure protection, emergency preparedness, and sustainable flood management strategies. Finally, the proposed and existing flood mitigation measures in the Red Sea drainage basins, including storage cisterns, culverts at road–wadi crossings, retardation barriers, and retardation dams (Figure 13), demonstrate practical approaches for reducing flood hazards and improving water resource management. Additionally, green infrastructure and constructed wetlands may further reduce runoff intensity and enhance water storage capacity.

5.4 Limitations and future research directions

Despite the strong predictive performance achieved by the hybrid AHP–ML models, several methodological limitations and sources of uncertainty should be acknowledged. The relatively high accuracy and AUC values may be influenced by the limited size of the flood inventory dataset and the inherent complexity of the ensemble machine learning algorithms, which may increase the risk of overfitting. In addition, the use of a random 70/30 train–test split may introduce spatial autocorrelation between training and testing samples, as geographically proximate locations often share similar environmental and geomorphological characteristics. Furthermore, because both training and testing datasets were derived from the same Sentinel-1 SAR-based flood inventory and historical flood records, the validation results reflect internal consistency rather than fully independent predictive performance. Consequently, the reported performance metrics, particularly AUC, may be slightly optimistic. These limitations highlight the need for spatially independent validation strategies, such as spatial cross-validation or independent sub-basin testing, to improve model robustness and reduce spatial bias in future studies.

Another important limitation of the present study is the absence of real-time flood observations and detailed temporal flood dynamics. The current flood susceptibility mapping (FSM) framework identifies flood-prone areas but does not quantify flood depth, discharge, or flow velocity. Incorporating coupled hydrological–hydrodynamic modeling approaches in future research would enhance the ability to simulate flood intensity and two-dimensional flow behavior. Additionally, the present analysis does not fully account for intra-annual and inter-annual climatic variability. The use of higher temporal resolution runoff and climate datasets could improve understanding of meteorological controls on flood generation processes. Improving spatial data resolution and integrating additional environmental conditioning factors may further enhance predictive performance. Finally, future research should evaluate the transferability of the proposed hybrid AHP–ML framework across different geomorphological and climatic settings to assess its robustness and generalizability for large-scale flood hazard applications.

5.5 Environmental sensitivity and desertification dynamics

The Environmental Sensitivity and Desertification Index (ESDI) results highlight the spatial heterogeneity of land degradation risk across the Safaga–El Quseir region. The classification into five sensitivity zones revealed that very high sensitivity areas are concentrated in coastal sabkhas and intensively cultivated floodplains, whereas elevated plateaus and reclaimed lands exhibit very low sensitivity. This spatial pattern underscores the combined influence of soil salinity, aridity, and anthropogenic pressure in shaping desertification dynamics. These findings are consistent with regional studies that identified vegetation decline and soil degradation as primary drivers of desertification in North African arid landscapes (Bouzekri et al., 2023; Ghebrezgabher et al., 2019). Similar patterns were observed in Eritrea, where Landsat-based assessments revealed strong correlations between vegetation loss and desertification hotspots (Ghebrezgabher et al., 2019), and in Algeria, where sandy desertification was linked to both climatic stress and unsustainable land use (Bouzekri et al., 2023). The present study extends these insights by integrating vegetation indices (NDVI, SAVI), soil degradation proxies, and land-use intensity into a composite sensitivity framework, thereby providing a more holistic diagnostic of environmental vulnerability. The identification of desertification hotspots along the Safaga–El Quseir tourism corridor emphasizes the role of rapid urban expansion and infrastructure development in accelerating land degradation. This observation aligns with broader evidence that anthropogenic drivers, particularly settlement growth and agricultural intensification, exacerbate ecological fragility in arid coastal systems (Anwer et al., 2025; Yousef et al., 2020). Importantly, the overlay of ESDI with flood susceptibility maps demonstrates that areas of high desertification sensitivity often coincide with flood-prone zones, amplifying compound risk and socio-economic exposure. From a policy perspective, these results reinforce the need for integrated land management strategies that simultaneously address flood hazards and desertification risks. The FAO and IUCN (2023) provide a relevant framework, emphasizing vegetation recovery, soil stabilization, and sustainable land use as critical interventions. Applying these restoration standards to the identified high-sensitivity zones would enhance ecological resilience and reduce vulnerability to compound hazards.

5.6 Multi-hazard interactions and compound risk

The integration of flood susceptibility and desertification sensitivity layers revealed distinct compound risk zones across the Safaga–El Quseir region. The multi-hazard susceptibility map demonstrated that areas of high and very high compound risk are concentrated along coastal floodplains, sabkhas, and intensively cultivated lands, whereas elevated plateaus and geomorphically stable terrain exhibit very low susceptibility. This spatial convergence of hazards underscores the importance of adopting integrated risk assessment frameworks in arid coastal systems. The identification of overlapping flood and desertification hotspots highlights the potential for cascading impacts, where hydrological extremes exacerbate soil degradation and vegetation decline. Such compound risks have been increasingly recognized in global multi-hazard studies, which emphasize that single-hazard approaches underestimate the true vulnerability of socio-ecological systems (Mudashiru et al., 2021; Edamo et al., 2022; Kim et al., 2025; Pourghasemi et al., 2020). The present study extends this perspective by demonstrating that hydrological drivers (elevation, runoff volume, drainage density) and ecological indicators (NDVI, aridity index) jointly determine susceptibility in the Red Sea coastal corridor. From a socio-economic standpoint, the overlap of flood and desertification hazards along the Safaga–El Quseir tourism corridor is particularly critical. Rapid urban expansion, infrastructure development, and tourism growth in these areas amplify exposure to compound risks, threatening both livelihoods and ecosystem services. This finding aligns with the (United Nations Office for Disaster Risk Reduction (UNDRR), 2023), which stresses that multi-hazard interactions must be explicitly incorporated into disaster risk reduction strategies to safeguard vulnerable communities. Policy implications are clear: integrated land-use planning and ecosystem restoration are required to mitigate compound risks. The FAO and IUCN (2023), currently under review, provide actionable guidance for restoring vegetation cover, stabilizing soils, and enhancing ecological resilience in high-risk zones. Applying these standards to the identified compound hotspots would reduce vulnerability to both flooding and desertification, while supporting sustainable development goals (SDG 13 and SDG 15).

5.7 Scenario-based hazard evolution and restoration implications

The scenario-based analyses revealed that compound hazard zones in the Safaga–El Quseir region are highly sensitive to both climatic forcing and socio-economic trajectories. Under RCP 4.5, moderate increases in runoff variability and aridity produced a 7% expansion of high-risk zones, primarily in lower catchments and coastal floodplains. In contrast, RCP 8.5 resulted in a more pronounced shift, with 15% of the study area transitioning into high or very high susceptibility classes, reflecting intensified runoff concentration and vegetation decline. These findings are consistent with global projections of hazard amplification under higher emission pathways (van Vuuren et al., 2011; Völz and Hinkel, 2023). Socio-economic pathways further modulated hazard evolution. SSP2 (middle-of-the-road development) maintained moderate hazard expansion, whereas SSP3 (regional rivalry with weak governance) exacerbated compound risk, particularly along the Safaga–El Quseir tourism corridor. Under SSP5 (fossil-fuel intensive growth), rapid urban expansion and infrastructure development led to a 12% increase in high-risk zones, underscoring the vulnerability of reclaimed lands and sabkhas to both flooding and desertification. These results align with recent scenario-based urban flood studies that emphasize the combined influence of climate forcing and land-use change on hazard susceptibility (Fu et al., 2025; Feng et al., 2026; Liu et al., 2025; Xu et al., 2024). The integration of these projections with the FAO Standards of Practice to Guide Ecosystem Restoration (Food and Agriculture Organization of the United Nations (FAO), 2025) provides a policy-relevant framework for prioritizing restoration interventions. Areas projected to shift into high or very high susceptibility under RCP 8.5 and SSP3/SSP5 should be targeted for vegetation recovery, soil stabilization, and sustainable land management. Such interventions would not only mitigate hazard exposure but also enhance ecological resilience and support the achievement of SDG 13 (Climate Action) and SDG 15 (Life on Land). From a strategic perspective, scenario-based hazard evolution underscores the need for adaptive management that integrates climate projections, socio-economic drivers, and restoration standards. By aligning predictive modeling with international restoration guidelines, the present study bridges the gap between scientific diagnostics and actionable resilience planning, offering a replicable framework for other arid coastal regions facing compound hazards.

5.7.1 Validation of ESDI and multi-hazard outputs

While the flood susceptibility mapping component was validated using ROC–AUC and cross-validation, the newly introduced ESDI and multi-hazard maps initially lacked quantitative validation. In this revision, we incorporated confusion matrices, kappa statistics, and accuracy indices to strengthen methodological transparency. The ESDI validation demonstrated substantial agreement with FAO and UNCCD datasets (κ = 0.71), while the multi-hazard framework achieved high predictive skill (ROC–AUC = 0.90). We acknowledge that these validations rely on global datasets, which may introduce resolution mismatches and do not fully substitute for independent field validation. Nevertheless, the quantitative results confirm that the integrated framework provides reliable diagnostics of environmental sensitivity and compound flood–desertification risks. Future work should incorporate higher-resolution local datasets and spatially independent validation to further improve robustness.

Quantitative validation results confirm the robustness of the proposed framework. The ESDI validation (Figure 14) demonstrates substantial agreement with FAO and UNCCD datasets, while the multi-hazard susceptibility validation (Figure 15) highlights strong predictive reliability in identifying compound flood–desertification risk zones. Together, these results substantiate the methodological soundness and transferability of the integrated approach for environmental risk assessment in arid coastal systems.

Novelty and Advancement Over Prior Work: Compared to El-Rayes et al. (2023) and Rashwan et al. (2023), who applied conventional MCDA and ML approaches independently, our hybrid AHP–ML framework improves predictive accuracy while enhancing interpretability through SHAP analysis. Unlike Youssef et al. (2022), who focused solely on flood susceptibility under present-day conditions, our study integrates flood susceptibility with desertification sensitivity to produce multi-hazard maps. Furthermore, by incorporating RCP/SSP scenarios and aligning outputs with FAO ecosystem restoration standards, we provide actionable guidance for resilience planning, representing a substantial methodological advance over prior work.

6 Conclusion

This study developed a comprehensive and interpretable flood susceptibility mapping framework for the Safaga–El Quseir region in Egypt’s Red Sea Mountains by coupling the Analytical Hierarchy Process (AHP) with machine learning classifiers (RF, XGB, GBM). Eleven flood-conditioning variables representing topographic, climatic, hydrological, and terrain characteristics were integrated with historical flood inventories derived from Sentinel-1 SAR and Google Earth Engine datasets. Diagnostic analyses confirmed the absence of multicollinearity, while Pearson correlation and Information Gain Ratio (InGR) identified elevation, volume runoff, slope, TWI, and aspect as dominant predictors. Model evaluation demonstrated strong predictive skill across all hybrid approaches, with AHP–RF achieving the highest accuracy (AUC = 0.95). SHAP-based interpretation further validated the critical influence of elevation, runoff, and drainage density, enhancing transparency and model explainability.

Beyond single-hazard mapping, the study advanced environmental diagnostics by generating an Environmental Sensitivity and Desertification Index (ESDI) map, which revealed high sensitivity in coastal sabkhas and intensively cultivated floodplains, and low sensitivity in elevated plateaus. The integration of flood susceptibility with ESDI produced a multi-hazard susceptibility map, highlighting compound risk zones where hydrological extremes and land degradation converge. Scenario-based analyses under RCP 4.5/8.5 and SSP pathways demonstrated significant expansion of high-risk zones under intensified climate forcing and unsustainable socio-economic trajectories. These projections emphasize the vulnerability of reclaimed lands and tourism corridors to future hazard amplification. The methodological novelty of this work lies in its multi-layer integration of geospatial datasets, machine learning, SHAP interpretability, and scenario-based projections, which collectively provide a replicable framework for hazard assessment in arid and semi-arid regions. By aligning scenario outputs with the FAO and IUCN (2023), the study bridges scientific diagnostics with actionable resilience planning. Restoration priorities such as vegetation recovery, soil stabilization, and sustainable land management are most urgently needed in zones projected to transition into high susceptibility under RCP 8.5 and SSP3/SSP5. Overall, the findings demonstrate that coupling AHP with machine learning not only improves predictive accuracy but also enhances practical applicability for integrated risk management. The incorporation of desertification sensitivity and scenario-based hazard evolution ensures that the framework addresses both present and future vulnerabilities. This research contributes to evidence-based decision-making for flood mitigation, land-use planning, and ecosystem restoration, offering a scalable model for climate-resilient hazard management in arid coastal environments.

References

• Abdel MoneimA. A. (2005). Overview of the geomorphological and hydrogeological characteristics of the Eastern Desert of Egypt. Hydrogeol. J.13, 416–425. 10.1007/s10040-004-0364y

• AbdelgawadA. G.HelalE.SobeihM. F.ElsayedH. (2024). Flood hazard mapping using a GIS-based morphometric analysis approach in arid regions, a case study in the Red Sea Region, Egypt. Appl. Water Sci.14 (4), 81. 10.1007/s13201-024-02130-5

• AbdelkareemM.MansourA. M. (2023). Risk assessment and management of vulnerable areas to flash flood hazards in arid regions using remote sensing and GIS-based knowledge-driven techniques. Nat. Hazards117 (3), 2269–2295. 10.1007/s11069-023-05942-x

• AbdraboK. I.KantoushS. A.SaberM.SumiT.HabibaO. M.ElleithyD.et al (2020). Integrated methodology for urban flood risk mapping at the microscale in ungauged regions: a case study of Hurghada, Egypt. Remote Sens.12 (21), 3548. 10.3390/rs12213548

• AbdraboK. I.KantoushS. A.EsmaielA.SaberM.SumiT.AlmamariM.et al (2023). An integrated indicator-based approach for constructing an urban flood vulnerability index as an urban decision-making tool using the PCA and AHP techniques: a case study of Alexandria, Egypt. Urban Clim.48, 101426. 10.1016/j.uclim.2023.101426

• AbdullahM. F.SirajS.HodgettR. E. (2021). An overview of multi-criteria decision analysis (MCDA) application in managing water-related disaster events: analyzing 20 years of literature for flood and drought events. Water13 (10), 1358. 10.3390/w13101358

• AbediR.CostacheR.Shafizadeh-MoghadamH.PhamQ. B. (2022). Flash-flood susceptibility mapping based on XGBoost, random forest and boosted regression trees. Geocarto Int.37 (19), 5479–5496. 10.1080/10106049.2021.1920636

• Abedi GheshlaghiH.FeizizadehB. (2021). GIS-based ensemble modelling of fuzzy system and bivariate statistics as a tool to improve the accuracy of landslide susceptibility mapping. Nat. Hazards107 (2), 1981–2014. 10.1007/s11069-021-04673-1

• Abu El-MagdS. A. (2022). Random forest and naïve Bayes approaches as tools for flash flood hazard susceptibility prediction, South Ras El-Zait, Gulf of Suez Coast, Egypt. Arabian J. Geosciences15 (3), 217. 10.1007/s12517-022-09531-3

• AbuziedS. M.MansourB. M. (2019). Geospatial hazard modeling for the delineation of flash flood-prone zones in Wadi Dahab basin, Egypt. J. Hydroinformatics21 (1), 180–206. 10.2166/hydro.2018.043

• AdeyemiA. B.KomolafeA. A.NakalembeC. L.IsmailaR. O.AdebayoA. D.BabayemiO. E. (2024). Integrated GIS-based MCDA and Machine Learning Techniques in Flood Susceptibility Mapping in Ala River Basin, Nigeria.

• AkinerM. E.GhasriM. (2024). Comparative assessment of deep belief network and hybrid adaptive neuro-fuzzy inference system model based on a meta-heuristic optimization algorithm for precise predictions of the potential evapotranspiration. Environ. Sci. Pollut. Res.31 (30), 42719–42749. 10.1007/s11356-024-33987-3

• Al-JuaidiA. E. (2023). The interaction of topographic slope with various geo-environmental flood-causing factors on flood prediction and susceptibility mapping. Environ. Sci. Pollut. Res.30 (21), 59327–59348. 10.1007/s11356-023-26616-y

• AliM. S.AbdelghanyM. K.MoftahH. M. (2023). Flood Susceptibility Assessment on the Proposed high-speed Train Track Qena-Hurghada-Safaga (QHS), Egypt.

• AliM.TahaM.AzizM. S.AhmedH.AhmedH. (2024). Flash flood prediction of Panjkora River, KPK, using Artificial Neural Networks (ANN) and Support Vector Machine (SVM). Tech. J.3 (ICACEE), 758–769. Available online at: https://tj.uettaxila.edu.pk/index.php/technical-journal/article/view/2068.

• AlsharhanA. S.KendallC. S. C. (2003). Holocene coastal carbonates and evaporites of the southern Arabian Gulf and their ancient analogues. Earth-Science Rev.61 (3-4), 191–243. 10.1016/s0012-8252(02)00110-1

• AmiriA.SoltaniK.EbtehajI.BonakdariH. (2024). A novel machine learning tool for current and future flood susceptibility mapping by integrating remote sensing and geographic information systems. J. Hydrology632, 130936. 10.1016/j.jhydrol.2024.130936

• AnwerH. A.MohamedT.HassanA. (2025). Assessing vegetation dynamics in Al Jazirah, Sudan using NDVI-based remote sensing techniques. J. Saudi Soc. Agric. Sci.24 (1), 18. 10.1007/s44447-025-00011-0

• ArnousM. O.AboulelaH. A.GreenD. R. (2011). Geo-environmental hazards assessment of the northwestern Gulf of Suez, Egypt. J. Coast. Conservation15, 37–50. 10.1007/s11852-010-0118-z

• AroraA.PandeyM.SiddiquiM. A.HongH.MishraV. N. (2021). Spatial flood susceptibility prediction in Middle Ganga Plain: comparison of frequency ratio and Shannon’s entropy models. Geocarto Int.36 (18), 2085–2116. 10.1080/10106049.2019.1687594

• AsiriM. M.AldehimG.AlruwaisN.AllafiR.AlzahraniI.NouriA. M.et al (2024). Coastal Flood risk assessment using ensemble multi-criteria decision-making with machine learning approaches. Environ. Res.245, 118042. 10.1016/j.envres.2023.118042

• AydinH. E.IbanM. C. (2023). Predicting and analyzing flood susceptibility using boosting-based ensemble machine learning algorithms with SHapley Additive exPlanations. Nat. Hazards116 (3), 2957–2991. 10.1007/s11069-022-05793-y

• BadawyM. (2008). Applications of Remote Sensing and Geographic Information Systems in Geomorphological Studies: Safaga-El Quseir Area, Red Sea Coast, Egypt as an Example. Doctoral dissertation, Mainz, Univ.

• BandS. S.JanizadehS.Chandra PalS.SahaA.ChakraborttyR.MelesseA. M.et al (2020). Flash flood susceptibility modeling using new approaches of hybrid and ensemble tree-based machine learning algorithms. Remote Sens.12 (21), 3568. 10.3390/rs12213568

• BarronT.HumeW. F. (1902). Topography and Geology of the Eastern Desert of Egypt, Central PortionNational Printing Department.

• BashirB.AlsalmanA. (2024). Morphometric characterization and dual analysis for flash flood hazard assessment of wadi Al-Lith watershed, Saudi Arabia. Water16 (22), 3333. 10.3390/w16223333

• BashirgonbadM.FarokhzadehB.GholamiV. (2024). Enhancing flood mapping through ensemble machine learning in the Gamasyab watershed, Western Iran. Environ. Sci. Pollut. Res.31 (38), 50427–50442. 10.1007/s11356-024-34501-5

• BeaudrieC.CorbettC. J.LewandowskiT. A.MalloyT.ZhouX. (2020). Evaluating the application of decision analysis methods in simulated alternatives assessment case studies: potential benefits and challenges of using MCDA. Integr. Environmental Assessment Management17 (1), 27–41. 10.1002/ieam.4316

• BeraS.DasA.MazumderT. (2022). Evaluation of machine learning, information theory and multi-criteria decision analysis methods for flood susceptibility mapping under varying spatial scale of analyses. Remote Sens. Appl. Soc. Environ.25, 100686. 10.1016/j.rsase.2021.100686

• BosworthW.KhalilS. M.LigiM.StockliD. F.McClayK. R. (2019). “Geology of Egypt: the northern red sea,” in The Geology of Egypt (Cham: Springer International Publishing), 343–374.

• BouzekriA.AlexandridisT. K.ToufikA.RebouhN. Y.ChenchouniH.KucherD.et al (2023). Assessment of the spatial dynamics of sandy desertification using remote sensing in Nemamcha region (Algeria). Egypt. J. Remote Sens. Space Sci.26 (3), 642–653. 10.1016/j.ejrs.2023.07.006

• BreimanL. (2001). Random forests. Mach. Learning45 (1), 5–32. 10.1023/a:1010933404324

• BrunnerM. I.SlaterL.TallaksenL. M.ClarkM. (2021). Challenges in modeling and predicting floods and droughts: a review. Wiley Interdiscip. Rev. Water8 (3), e1520. 10.1002/wat2.1520

• BryndalT.FranczakP.KroczakR.CabajW.KołodziejA. (2017). The impact of extreme runoff and flash floods on the flood risk management process and geomorphological changes in small Carpathian catchments: a case study of the Kasiniczanka river (Outer Carpathians, Poland). Nat. Hazards88, 95–120. 10.1007/s11069-017-2858-7

• BullL. J.KirkbyM. J.ShannonJ.HookeJ. M. (2000). The impact of rainstorms on floods in ephemeral channels in southeast Spain. Catena38 (3), 191–209. 10.1016/s0341-8162(99)00071-5

• ChakrabortyS.MukhopadhyayS. (2019). Assessing flood risk using analytical hierarchy process (AHP) and geographical information system (GIS): application in Coochbehar district of West Bengal, India. Nat. Hazards99, 247–274. 10.1007/s11069-019-03737-7

• ChiarellaD.ScarselliN.AdamJ. (2024). Regional Geology and Tectonics: Phanerozoic Rift Systems and Sedimentary Basins (Elsevier), 2.

• CostacheR.PhamQ. B.SharifiE.LinhN. T. T.AbbaS. I.VojtekM.et al (2019). Flash-flood susceptibility assessment using multi-criteria decision making and machine learning supported by remote sensing and GIS techniques. Remote Sens.12 (1), 106. 10.3390/rs12010106

• CostacheR.ArabameriA.ElkhrachyI.GhorbanzadehO.PhamQ. M. (2021). Detection of areas prone to flood risk using state-of-the-art machine learning models. Geomat. Nat. Hazards Risk12 (1), 1488–1507. 10.1080/19475705.2021.1920480

• DahriN.YousfiR.BouamraneA.AbidaH.PhamQ. B.DerdousO. (2022). Comparison of analytic network process and artificial neural network models for flash flood susceptibility assessment. J. Afr. Earth Sci.193, 104576. 10.1016/j.jafrearsci.2022.104576

• DienoW. U.PathanS. A. (2025). Flood susceptibility mapping and 3D flood simulation in the Jia Bharali River Basin, Sonitpur district, India, using the Analytical Hierarchy Process (AHP). Remote Sens. Earth Syst. Sci.8, 1–31. 10.1007/s41976-025-00226-3

• DingL.MaL.LiL.LiuC.LiN.YangZ.et al (2021). A survey of remote sensing and geographic information system applications for flash floods. Remote Sens.13 (9), 1818. 10.3390/rs13091818

• DouglasE. M.VogelR. M.KrollC. N. (2000). Trends in floods and low flows in the United States: impact of spatial correlation. J. Hydrology240 (1-2), 90–105. 10.1016/s0022-1694(00)00336-x

• DuanL.LiuC.XuH.PanH.LiuH.YanX.et al (2022). Susceptibility assessment of flash floods: a bibliometrics analysis and review. Remote Sens.14 (21), 5432. 10.3390/rs14215432

• EdamoM. L.UkumoT. Y.LohaniT. K.AyanaM. T.AyeleM. A.MadaZ. M.et al (2022). A comparative assessment of multi-criteria decision-making analysis and machine learning methods for flood susceptibility mapping and socio-economic impacts on flood risk in Abela-Abaya floodplain of Ethiopia. Environ. Challenges9, 100629. 10.1016/j.envc.2022.100629

• EfraimidouE.SpiliotisM. (2024). A gis-based flood risk assessment using the decision-making trial and evaluation laboratory approach at a regional scale. Environ. Process.11 (1), 9. 10.1007/s40710-024-00683-w

• El BastawesyM.AttwaM.HafeezT. H. A.GadA. (2019). Flash floods and groundwater evaluation for the non-gauged dryland catchment using remote sensing, GIS and DC resistivity data: a case study from the Eastern Desert of Egypt. J. Afr. Earth Sci.152, 245–255. 10.1016/j.jafrearsci.2019.02.004

• El GabyS.ListF. K.TehraniR. (2017). “The basement complex of the Eastern Desert and Sinai,” in The Geology of Egypt (Routledge), 175–184.

• El RamlyI. M. (2010). The calanchio–uweinat uplift and its impact on the groundwater movement between the kufra basin (libya) and the dakhla basin (egypt). J. Egypt. Geophys. Soc.8 (1), 213–233. 10.21608/jegs.2010.380914

• El-BagouryH.DarwishM. H.HassanS. H.OhS. E.AttiaK. A.MegahedH. A. (2025). Potentiality delineation of groundwater recharge in arid regions using multi-criteria analysis. Water17 (5), 766. 10.3390/w17050766

• El-RawyM.ElsadekW. M.De SmedtF. (2023). Flood hazard assessment and mitigation using a multi-criteria approach in the Sinai Peninsula, Egypt. Nat. Hazards115 (1), 215–236. 10.1007/s11069-022-05551-0

• El-RayesA. E.ArnousM. O.HelmyA. M. (2023). GIS-based flash flooding susceptibility analysis and water management in arid mountain ranges: Safaga Region, Red Sea Mountains, Egypt. J. Mt. Sci.20 (12), 3665–3686. 10.1007/s11629-023-8142-2

• EmbabiN. S. (2018). Landscapes and Landforms of Egypt. Springer.

• FAOS.IUCNC. (2023). Standards of practice to guide ecosystem restoration. A contribution to the United Nations Decade on Ecosystem Restoration. Summary report, 1–17.

• FeizbahrM.BrakeN.ArbabkhahH.Hariri AsliH.WoodsK. (2025). Flood susceptibility mapping using machine learning and Geospatial-Sentinel-1 SAR integration for enhanced early warning systems. Remote Sens.17 (20), 3471. 10.3390/rs17203471

• FeloniE.AnayiotosA.BaltasE. (2022). A spatial analysis approach for urban flood occurrence and flood impact based on geomorphological, meteorological, and hydrological factors. Geographies2 (3), 516–527. 10.3390/geographies2030031

• FengW.LiuY.ZhuA.MaoJ.SunT.XuQ.et al (2026). Modeling urban planning contributions to flood resilience under shared socioeconomic pathways. Npj Urban Sustain.6, 45. 10.1038/s42949-026-00353-w

• FernándezJ.AgirreI.Perez-CerrolazaJ.BelategiL.AdellA.DonzellaC.et al (2024). “AI-FSM: towards functional safety management for artificial intelligence-based critical systems,” in CARS@ EDCC2024 Workshop-Critical Automotive Applications: Robustness and Safety. Available online at: https://www.linkedin.com/pulse/ai-fsm-towards-functional-safety-management-critical-fern%C3%A1ndez-mu%C3%B1oz-pqgdf?trk=public_post_main-feed-card_reshare_feed-article-content.

• FilianotiP.GurnariL.ZemaD. A.BombinoG.SinagraM.TucciarelliT. (2020). An evaluation matrix to compare computer hydrological models for flood predictions. Hydrology7 (3), 42. 10.3390/hydrology7030042

• Food and Agriculture Organization of the United Nations (FAO) (2025). Ecosystem restoration monitoring. Rome, Italy: FAO. Available online at: https://www.fao.org/ecosystem-restoration-monitoring/en

• FuX.XueF.LiuY.ChenF.YangH. (2025). Evaluation of urban flood susceptibility under the influence of urbanization based on shared socioeconomic pathways. Land14 (3), 621. 10.3390/land14030621

• GhebrezgabherM.YangT.YangX.WangC. (2019). Assessment of desertification in Eritrea: land degradation based on Landsat images. J. Arid Land11 (3), 319–331. 10.1007/s40333-019-0096-4

• HamimiZ.Abd El-WahedM. A. (2020). “Suture(s) and major shear zones in the neoproterozoic basement of egypt,” in the Geology of Egypt.Editors Hamimi, Z., El-Barkooky, A., Martínez Frías, J., Fritz, H., Abd El-Rahman, Y (Springer, Cham: Regional Geology Reviews)153–189. 10.1007/978-3-030-15265-9_5

• HasanM. M.TalhaM.AkterM. M.FerdousM. T.MojumderP.RoyS. K.et al (2025). Assessing the performance of machine learning and analytical hierarchy process (AHP) models for rainwater harvesting potential zone identification in hilly region, Bangladesh. J. Asian Earth Sci.X (13), 100189. 10.1016/j.jaesx.2024.100189

• HeF.LiuS.MoX.WangZ. (2025). Interpretable flash flood susceptibility mapping in Yarlung Tsangpo River Basin using H2O Auto-ML. Sci. Rep.15 (1), 1702. 10.1038/s41598-024-84655-y

• HerreraP. A.MarazuelaM. A.HofmannT. (2022). Parameter estimation and uncertainty analysis in hydrological modeling. Wiley Interdiscip. Rev. Water9 (1), e1569. 10.1002/wat2.1569

• HussainM.TayyabM.UllahK.UllahS.RahmanZ. U.ZhangJ.et al (2023). Development of a new integrated flood resilience model using machine learning with GIS-based multi-criteria decision analysis. Urban Clim.50, 101589. 10.1016/j.uclim.2023.101589

• IslamA. R. M. T.BappiM. M. R.AlqadhiS.BindajamA. A.MallickJ.TalukdarS. (2023). Improvement of flood susceptibility mapping by introducing hybrid ensemble learning algorithms and high-resolution satellite imageries. Nat. Hazards119 (1), 1–37. 10.1007/s11069-023-06106-7

• IslamT.ZelekeE. B.AfrozM.MelesseA. M. (2025). A systematic review of urban flood susceptibility mapping: remote sensing, machine learning, and other modeling approaches. Remote Sens.17 (3), 524. 10.3390/rs17030524

• IUSS Working Group (2015). World reference base for soil resources 2014, update 2015 international soil classification system for naming soils and creating legends for soil maps. FAO, Rome: World Soil Resources Reports No. 106.

• JonkmanS. N.CurranA.BouwerL. M. (2024). Floods have become less deadly: an analysis of global flood fatalities 1975–2022. Nat. Hazards120 (7), 6327–6342. 10.1007/s11069-024-06444-0

• KhanT. A.AlamM. M.ShahidZ.Su’UdM. M. (2020). Investigation of flash floods on early basis: a factual comprehensive review. IEEE Access8, 19364–19380. 10.1109/access.2020.2967496

• KhattabM. I.FadlM. E.MegahedH. A.SaleemA. M.El-SaadawyO.DrososM.et al (2025). Evaluation of the soil conservation service curve number (SCS-CN) method for flash flood runoff estimation in arid regions: a case Study of central Eastern desert, Egypt. Hydrology12 (3), 54. 10.3390/hydrology12030054

• KhodaeiH.Nasiri SalehF.Nobakht DalirA.ZareiE. (2025). Future flood susceptibility mapping under climate and land use change. Sci. Rep.15, 12394. 10.1038/s41598-025-97008-0

• KimC.ParkS.HanH. (2025). Multi-hazard susceptibility mapping using machine learning approaches: a case study of South Korea. Remote Sens.17 (10), 1660. 10.3390/rs17101660

• Kocur-BeraK.MałekA. (2024). Assessing the feasibility of using remote sensing data and vegetation indices in the estimation of land subject to consolidation. Sensors24 (23), 7736. 10.3390/s24237736

• KotbA.TahaA. I.ElnazerA. A.BasheerA. A. (2024). Global insights on flood risk mitigation in arid regions using geomorphological and geophysical modeling from a local case study. Sci. Rep.14 (1), 19975. 10.1038/s41598-024-69541-x

• KumarR. (2025). A comprehensive review of MCDM methods, applications, and emerging trends. Decis. Mak. Adv.3 (1), 185–199. 10.31181/dma31202569

• KuruvilaR. (2024). AHP-based flood susceptibility and evacuation route mapping in the Ernakulam District of Kerala. PhD dissertation. Pondicherry University, Puducherry, India.

• LatifR. M. A.HeJ. (2024). Flood susceptibility mapping in Punjab, Pakistan: a hybrid approach integrating remote sensing and analytical hierarchy process. Atmosphere16 (1), 22. 10.3390/atmos16010022

• LealM.ReisE.SantosP. P. (2022). Exploring spatial relationships between stream channel features, water depths and flow velocities during flash floods using HEC-GeoRAS and Geographic information systems. J. Geogr. Sci.32 (4), 757–782. 10.1007/s11442-022-1971-z

• LevyJ. K. (2005). Multiple criteria decision making and decision support systems for flood risk management. Stoch. Environ. Res. Risk Assess.19 (6), 438–447. 10.1007/s00477-005-0009-2

• LiaoH.HeY.WuX.WuZ.BausysR. (2023). Reimagining multi-criterion decision making by data-driven methods based on machine learning: a literature review. Inf. Fusion100, 101970. 10.1016/j.inffus.2023.101970

• LinK.ChenH.XuC. Y.YanP.LanT.LiuZ.et al (2020). Assessment of flash flood risk based on improved analytic hierarchy process method and integrated maximum likelihood clustering algorithm. J. Hydrology584, 124696. 10.1016/j.jhydrol.2020.124696

• LiuQ.LiuG.HuangC. (2018). Monitoring desertification processes in the Mongolian Plateau using MODIS tasseled cap transformation and TGSI time series. J. Arid Land10 (1), 12–26. 10.1007/s40333-017-0109-0

• LiuJ.XiongJ.ChenY.SunH.ZhaoX.TuF.et al (2023). A new avenue to improve the performance of integrated modeling for flash flood susceptibility assessment: applying cluster algorithms. Ecol. Indic.146, 109785. 10.1016/j.ecolind.2022.109785

• LiuW.ZhangX.FengQ.EngelB. A. (2025). City-scale integrated flood risk prediction under future climate change and urbanization based on the shared socioeconomic pathways (SSP) scenarios. J. Hydrology603, 132971. 10.1016/j.jhydrol.2025.132971

• LuY.ZhaiG.ZhaiW. (2024). Quantifying urban spatial resilience using multi-criteria decision analysis (MCDA) and back propagation neural network (BPNN). Int. J. Disaster Risk Reduct.111, 104694. 10.1016/j.ijdrr.2024.104694

• LundbergS.LeeS. I. (2017). “A unified approach to interpreting model predictions,” in Advances in Neural Information Processing Systems. 10.48550/arXiv.1705.07874

• MaM.ZhaoG.HeB.LiQ.DongH.WangS.et al (2021). XGBoost-based method for flash flood risk assessment. J. Hydrology598, 126382. 10.1016/j.jhydrol.2021.126382

• MahmoudS. H.GanT. Y. (2018). Multi-criteria approach to develop flood susceptibility maps in arid regions of Middle East. J. Cleaner Production196, 216–229. 10.1016/j.jclepro.2018.06.047

• MajeedM.LuL.AnwarM. M.TariqA.QinS.El-HefnawyM. E.et al (2023). Prediction of flash flood susceptibility using integrating analytic hierarchy process (AHP) and frequency ratio (FR) algorithms. Front. Environ. Sci.10, 1037547. 10.3389/fenvs.2022.1037547

• ManandharB.CuiS.WangL.ShresthaS. (2023). Urban flood hazard assessment and management practices in south Asia: a review. Land12 (3), 627. 10.3390/land12030627

• MashalyJ.GhoneimE. (2018). Flash flood hazard using optical, radar, and stereo-pair derived dem: eastern desert, Egypt. Remote Sens.10 (8), 1204. 10.3390/rs10081204

• MegahedH. A.AbdoA. M.AbdelRahmanM. A.ScopaA.HegazyM. N. (2023a). Frequency ratio model as tools for flood susceptibility mapping in urbanized areas: a case study from Egypt. Appl. Sci.13 (16), 9445. 10.3390/app13169445

• MegahedH. A.GabAllahH. M.RamadanR. H.AbdelRahmanM. A.D’AntonioP.ScopaA.et al (2023b). Groundwater quality assessment using multi-criteria GIS modeling in drylands: a case study at El-Farafra Oasis, Egyptian Western Desert. Water15 (7), 1376. 10.3390/w15071376

• MegahedH. A.FarragA. E. H. A.MohamedA. A.D’AntonioP.ScopaA.AbdelRahmanM. A. (2023c). Groundwater recharge potentiality mapping in Wadi Qena, Eastern Desert basins of Egypt for sustainable agriculture base using geomatics approaches. Hydrology10 (12), 237. 10.3390/hydrology10120237

• MegahedH. A.FarragA. E. H. A.GabAllahH. M.AbdelRahmanM. A.BadawyR. M. (2024). Develop of a machine learning model to evaluate the hazards of sand dunes. Earth Sci. Inf.17 (5), 4001–4025. 10.1007/s12145-024-01284-9

• MegahedH. A.FarragA. E. H. A.MohamedA. A.DarwishM. H.AbdelRahmanM. A.El-BagouryH.et al (2025). GIS-based modeling and analytical approaches for groundwater quality suitability for different purposes in the Egyptian Nile Valley, a case study in Wadi Qena. Front. Water7, 1502169. 10.3389/frwa.2025.1502169

• MirzaeiS.VafakhahM.PradhanB.AlaviS. J. (2021). Flood susceptibility assessment using extreme gradient boosting (EGB), Iran. Earth Sci. Inf.14 (1), 51–67. 10.1007/s12145-020-00530-0

• MishraP.PrasadS. K. (2024). Reviewing flood susceptibility mapping utilizing machine learning: insights and prospects for future research. In 2024 3rd International Conference on Applied Artificial Intelligence and Computing (ICAAIC) (649–655). IEEE.

• MohamedM. A. M. (2023).Rifting Dynamics and Margin Architecture in Northern Red Sea, Egypt (Doctoral Dissertation, Khalifa University of Science).

• MohamedH. I.AbozeidG.ShehataS.MohamedE. (2021). Hydrological parameters affecting flash floods: case study Wadi Al-Baroud El-Abiad, Safaga City, red sea governorate. Aswan Univ. J. Environ. Stud.2 (1), 42–56. 10.21608/aujes.2021.149766

• MooreI. D.BurchG. J. (1986). Physical basis of the length‐slope factor in the universal soil loss equation. Soil Sci. Soc. Am. J.50 (5), 1294–1298. 10.2136/sssaj1986.03615995005000050042x

• MoustafaA. R.KhalilS. M. (2019). “Structural setting and tectonic evolution of the Gulf of Suez, NW Red Sea and Gulf of Aqaba rift systems,” in The Geology of Egypt (Cham: Springer International Publishing), 295–342.

• MudashiruR. B.SabtuN.AbustanI.BalogunW. (2021). Flood hazard mapping methods: a review. J. Hydrology603, 126846. 10.1016/j.jhydrol.2021.126846

• MudashiruR. B.SabtuN.AbdullahR.SalehA.AbustanI. (2022). Optimality of flood influencing factors for flood hazard mapping: an evaluation of two multi-criteria decision-making methods. J. Hydrology612, 128055. 10.1016/j.jhydrol.2022.128055

• MuellerT.SeginA.WeigandC.SchmittR. H. (2023). Feature selection for measurement models. Int. Journal Quality and Reliability Management40 (3), 777–800. 10.1108/ijqrm-07-2021-0245

• MunirB. A.AhmadS. R.HafeezS. (2019). Integrated hazard modeling for simulating torrential stream response to flash flood events. ISPRS International Journal Geo-Information9 (1), 1. 10.3390/ijgi9010001

• NachappaT. G.PiralilouS. T.GholamniaK.GhorbanzadehO.RahmatiO.BlaschkeT. (2020). Flood susceptibility mapping with machine learning, multi-criteria decision analysis and ensemble using Dempster Shafer Theory. J. Hydrology590, 125275. 10.1016/j.jhydrol.2020.125275

• NguyenT. T.HoangA. H.PhamT. t. h.TranT. t. t. (2023). Flash flood hazard mapping using landsat-8 imagery, Ahp, and Gis in the Ngan Sau and Ngan Pho River basins, North-Central Vietnam. Geogr. Environ. Sustain16 (2), 57–67. 10.24057/2071-9388-2022-117

• NguyenD.ChouT.HoangT.ChenM. (2023). Flood susceptibility mapping using machine learning algorithms: a case study in Huong Khe District, Ha Tinh Province, Vietnam. Int. J. Geoinformatics19 (7), 1–15. 10.52939/ijg.v19i7.2739

• NkekiF. N.BelloE. I.AgbajeI. G. (2022). Flood risk mapping and urban infrastructural susceptibility assessment using a GIS and analytic hierarchical raster fusion approach in the Ona River Basin, Nigeria. Int. J. Disaster Risk Reduct.77, 103097. 10.1016/j.ijdrr.2022.103097

• NoharaY.MatsumotoK.SoejimaH.NakashimaN. (2022). Explanation of machine learning models using shapley additive explanation and application for real data in hospital. Comput. Methods Programs Biomed.214, 106584. 10.1016/j.cmpb.2021.106584

• NoorunnaharM.ChowdhuryA. H.MilaF. A. (2023). A tree based eXtreme Gradient Boosting (XGBoost) machine learning model to forecast the annual rice production in Bangladesh. PloS One18 (3), e0283452. 10.1371/journal.pone.0283452

• NurjahanS.RahmanM. M.AhmedI.Al-HabibM. J. (2026). Spatiotemporal analysis of flood susceptibility mapping in Asia: a comprehensive review of methodological and parameter-based approaches. Nat. Hazards Rev.27 (3), 03126002. 10.1061/nhrefo.nheng-2737

• OthmanA.El-SaoudW. A.HabeebullahT.ShaabanF.AbotalibA. Z. (2023). Risk assessment of flash flood and soil erosion impacts on electrical infrastructures in overcrowded mountainous urban areas under climate change. Reliab. Eng. and Syst. Saf.236, 109302. 10.1016/j.ress.2023.109302

• OtsuN. (1979). Threshold selection method from gray-level histograms. IEEE Trans. Syst. Man. Cybern.9, 62–66. 10.1109/tsmc.1979.4310076

• PathakL. (2026). A multi-model statistical, machine learning, and deep learning framework for landslide susceptibility in Nepal’s mid-hills. Discov. Hazards2, 20. 10.1007/s44475-026-00029-0

• PenkiR.BasinaS. S.TanniruS. R. (2023). Application of geographical information system-based analytical hierarchy process modeling for flood susceptibility mapping of Krishna District in Andhra Pradesh. Environ. Sci. Pollut. Res.30 (44), 99062–99075. 10.1007/s11356-022-22924-x

• PhamB. T.AvandM.JanizadehS.PhongT. V.Al-AnsariN.HoL. S.et al (2020). GIS based hybrid computational approaches for flash flood susceptibility assessment. Water12 (3), 683. 10.3390/w12030683

• PourghasemiH. R.KariminejadN.AmiriM.EdalatM.ZarafsharM.BlaschkeT.et al (2020). Assessing and mapping multi-hazard risk susceptibility using a machine learning technique. Sci. Rep.10, 3203. 10.1038/s41598-020-60191-3

• RamadanE. M.ShahinH. A.Abd-ElhamidH. F.ZelenakovaM.EldeebH. M. (2022). Evaluation and mitigation of flash flood risks in arid regions: a case study of Wadi Sudr in Egypt. Water14 (19), 2945. 10.3390/w14192945

• RashwanM.MohamedA. K.AlshehriF.AlmadaniS.KhattabM.MohamedL. (2023). Flash flood hazard assessment along the red sea Coast using remote sensing and GIS techniques. ISPRS Int. J. Geo-Information12 (11), 465. 10.3390/ijgi12110465

• RawatP. K.PantC. C.TiwariP. C.PantP. D.SharmaA. K. (2012). Spatial variability assessment of river‐line floods and flash floods in Himalaya: a case study using GIS. Disaster Prev. Manag. An Int. J.21 (2), 135–159. 10.1108/09653561211219955

• RezvaniH.ArfaA.Shafizadeh-MoghadamM.MinaeiM. (2026). An XGBoost-SHAP framework for interpretable and probabilistic flood susceptibility mapping. Nat. Hazards122 (6), 225. 10.1007/s11069-025-07908-7

• RiazM. T.RiazM. T.RehmanA.BindajamA. A.MallickJ.AbdoH. G. (2024). An integrated approach of support vector machine (SVM) and weight of evidence (WOE) techniques to map groundwater potential and assess water quality. Sci. Rep.14 (1), 26186. 10.1038/s41598-024-76607-3

• RoyS.BoseA.ChowdhuryI. R. (2021). Flood risk assessment using geospatial data and multi-criteria decision approach: a study from historically active flood-prone region of Himalayan foothill, India. Arabian J. Geosciences14 (11), 999. 10.1007/s12517-021-07324-8

• SaadM. S. H.AliM. I.RaziP. Z.RamliN. I.JayaR. P. (2024). Exploring the factors and impacts of flash flood vulnerability in various areas of Malaysia: a content analysis. Disaster Civ. Eng. Archit.1 (1), 55–82. 10.70028/dcea.v1i1.11

• SaatyT. L. (1980). The Analytical Hierarchy Process. New York: McGraw-Hill.

• SahlaouiH.NayyarA.AgoujilS.JaberM. M. (2021). Predicting and interpreting student performance using ensemble models and shapley additive explanations. IEEe Access9, 152688–152703. 10.1109/access.2021.3124270

• SahooS. K.GoswamiS. S. (2023). A comprehensive review of multiple criteria decision-making (MCDM) methods: advancements, applications, and future directions. Decis. Making Advances1 (1), 25–48. 10.31181/dma1120237

• SalmaNikhilS.DanumahJ. H.PrasadM. K.NazarN.SahaS.et al (2023). Prediction capability of the MCDA-AHP model in wildfire risk zonation of a protected area in the Southern Western Ghats. Environ. Sustain.6 (1), 59–72. 10.1007/s42398-022-00259-0

• SchmidtL.HeßeF.AttingerS.KumarR. (2020). Challenges in applying machine learning models for hydrological inference: a case study for flooding events across Germany. Water Resources Research56 (5), e2019WR025924. 10.1029/2019wr025924

• SebokaG. N. (2016). Spatial assessment of NDVI as an indicator of desertification in ethiopia using remote sensing and GIS (master’s Thesis). Lund University, Department of Physical Geography and Ecosystem Science. Available online at: https://web.archive.org/web/20180512023416id_/http://lup.lub.lu.se/luur/download?func=downloadFile&recordOId=8877776&fileOId=8877790.

• ShafikW. (2025). “Machine learning techniques for multicriteria decision-making,” in Multi-Criteria Decision-Making and Optimum Design with Machine Learning (CRC Press), 165–194. 10.1201/9781032635170-13

• ShahS. A.AiS. (2024). Flood susceptibility mapping contributes to disaster risk reduction: a case study in Sindh, Pakistan. Int. J. Disaster Risk Reduct.108, 104503. 10.1016/j.ijdrr.2024.104503

• ShawkyM.HassanQ. K. (2023). Geospatial modeling based-multi-criteria decision-making for flash flood susceptibility zonation in an arid area. Remote Sens.15 (10), 2561. 10.3390/rs15102561

• ShehadehA.AlshboulO.Al MamlookR. E.HamedatO. (2021). Machine learning models for predicting the residual value of heavy construction equipment: an evaluation of modified decision tree, LightGBM, and XGBoost regression. Automation Constr.129, 103827. 10.1016/j.autcon.2021.103827

• SiauK.RossiM. (2011). Evaluation techniques for systems analysis and design modelling methods–a review and comparative analysis. Inf. Syst. J.21 (3), 249–268. 10.1111/j.1365-2575.2007.00255.x

• SibindiR.MwangiR. W.WaitituA. G. (2023). A boosting ensemble learning based hybrid light gradient boosting machine and extreme gradient boosting model for predicting house prices. Eng. Rep.5 (4), e12599. 10.1002/eng2.12599

• SpanidisP. M.RoumposC.PavloudakisF. (2021). A fuzzy-AHP methodology for planning the risk management of Natural hazards in surface mining projects. Sustainability13 (4), 2369. 10.3390/su13042369

• SubraeluP.AhmedA.EbraheemA. A.SherifM.MirzaS. B.RidouaneF. L.et al (2023). Risk assessment and mapping of flash flood vulnerable zones in arid region, fujairah city, uae-using remote sensing and gis-based analysis. Water15 (15), 2802. 10.3390/w15152802

• TaherizadehM.NiknamA.Nguyen-HuyT.MezősiG.SarliR. (2023). Flash flood-risk areas zoning using integration of decision-making trial and evaluation laboratory, GIS-based analytic network process and satellite-derived information. Nat. Hazards118 (3), 2309–2335. 10.1007/s11069-023-06089-5

• TalukdarS.GhoseB.SalamR.MahatoS.PhamQ. B.AvandM.et al (2020). Flood susceptibility modeling in Teesta River basin, Bangladesh using novel ensembles of bagging algorithms. Stoch. Environ. Res. Risk Assess.34 (12), 2277–2300. 10.1007/s00477-020-01862-5

• TanX. Z.LiY.WuX. X.DaiC.ZhangX. L.CaiY. P. (2024). Identification of the key driving factors of flash flood based on different feature selection techniques coupled with random forest method. J. Hydrology Regional Stud.51, 101624. 10.1016/j.ejrh.2023.101624

• TariqA.YanJ.GhaffarB.QinS.MousaB. G.SharifiA.et al (2022). Flash flood susceptibility assessment and zonation by integrating analytic hierarchy process and frequency ratio model with diverse spatial data. Water14 (19), 3069. 10.3390/w14193069

• United Nations Office for Disaster Risk Reduction (UNDRR) (2023). Report of the secretary-general on the implementation of the sendai framework for disaster risk reduction 2015–2030. Geneva, Switzerland: United Nations General Assembly, Report A/78/267.

• UruvilaR. (2024). AHP-based Flood Susceptibility and Evacuation Route Mapping in the Ernakulam District of Kerala. Doctoral dissertation, Pondicherry University. Available online at: https://urgent-project.net/uploads/page/files/RITHU%20KURUVILA%20DISSERTATION.pdf.

• Van VuurenD. P.EdmondsJ.KainumaM.RiahiK.ThomsonA.HibbardK.et al (2011). The representative concentration pathways: an overview. Clim. Change109 (1–2), 5–31. 10.1007/s10584-011-0148-z

• VölzV.HinkelJ. (2023). Climate learning scenarios for adaptation decision analyses: review and classification. Clim. Risk Manag.40, 100512. 10.1016/j.crm.2023.100512

• VujovićF.ValjarevićA.DurlevićU.MorarC.GramaV.SpalevićV.et al (2025). A comparison of the AHP and BWM models for the flash flood susceptibility assessment: a case study of the Ibar River Basin in Montenegro. Water17 (6), 844. 10.3390/w17060844

• WangY.TangQ.WrightN. (2022). “Managing urban flood risk and building resilience in a changing climate,” in Climate Risk and Sustainable Water Management. Editors TangQ.LengG. (Cambridge University Press), 315–341.

• WangM.LiY.YuanH.ZhouS.WangY.IkramR. M. A.et al (2023). An XGBoost-SHAP approach to quantifying morphological impact on urban flooding susceptibility. Ecol. Indic.156, 111137. 10.1016/j.ecolind.2023.111137

• WaqasH.LuL.TariqA.LiQ.BaqaM. F.XingJ.et al (2021). Flash flood susceptibility assessment and zonation using an integrating analytic hierarchy process and frequency ratio model for the Chitral District, Khyber Pakhtunkhwa, Pakistan. Water13 (12), 1650. 10.3390/w13121650

• WuZ.XueW.XuH.YanD.WangH.QiW. (2022). Urban flood risk assessment in Zhengzhou, China, based on a D-number-improved analytic hierarchy process and a self-organizing map algorithm. Remote Sens.14 (19), 4777. 10.3390/rs14194777

• XuK.HanZ.XuH.BinL. (2023). Rapid prediction model for urban floods based on a light gradient boosting machine approach and hydrological–hydraulic model. Int. J. Disaster Risk Sci.14 (1), 79–97. 10.1007/s13753-023-00465-2

• XuW.CaiX.YuQ.ProverbsD.XiaT. (2024). Modelling trends in urban flood resilience towards improving the adaptability of cities. Water16 (11), 1614. 10.3390/w16111614

• YangC.GuanX.XuQ.XingW.ChenX.ChenJ.et al (2024). How can SHAP (SHapley additive exPlanations) interpretations improve deep learning based urban cellular automata model?Comput. Environ. Urban Syst.111, 102133. 10.1016/j.compenvurbsys.2024.102133

• YinJ.GaoY.ChenR.YuD.WilbyR.WrightN.et al (2023). Flash floods: why are more of them devastating the world’s driest regions?Nature615 (7951), 212–215. 10.1038/d41586-023-00626-9

• YousefR. T.HassanI. T.AlkhaqaniA. (2020). Desertification detection by NDVI and other indices: a case study in Najaf. J. Xidian Univ.14 (9), 126. 10.37896/jxu14.9/126

• YoussefA. M.PourghasemiH. R.El-HaddadB. A. (2022). Advanced machine learning algorithms for flood susceptibility modeling—performance comparison: red Sea, Egypt. Environ. Sci. Pollut. Res.29 (44), 66768–66792. 10.1007/s11356-022-20213-1

• ZhranM.GhanemK.TariqA.AlshehriF.JinS.DasJ.et al (2024). Exploring a GIS-based analytic hierarchy process for spatial flood risk assessment in Egypt: a case study of the Damietta branch. Environ. Sci. Eur.36 (1), 1–25. 10.1186/s12302-024-01001-9