Drainage Network Generation for Urban Pluvial Flooding (UPF) Using Generative Adversarial Networks (GANs) and GIS Data

Abstract

Mapping urban pluvial flooding (UPF) in data-scarce regions poses significant challenges, particularly when drainage systems are inadequate or outdated. These limitations hinder effective flood mitigation and risk assessment. This study proposes an innovative approach to address these challenges by integrating deep learning (DL) models with traditional methods. First, deep convolutional generative adversarial networks (DCGANs) were employed to enhance drainage network data generation. Second, deep recurrent neural networks (DRNNs) and multi-criteria decision analysis (MCDA) methods were implemented to assess UPF. The study compared the performance of these approaches, highlighting the potential of DL models in providing more accurate and robust flood mapping outcomes. The methodology was applied to Lahore, Pakistan—a rapidly urbanizing and data-scarce region frequently impacted by UPF during monsoons. High-resolution ALOS PALSAR DEM data were utilized to extract natural drainage networks, while synthetic datasets generated by GANs addressed the lack of historical flood data. Results demonstrated the superiority of DL-based approaches over traditional MCDA methods, showcasing their potential for broader applicability in similar regions worldwide. This research emphasizes the role of DL models in advancing urban flood mapping, providing valuable insights for urban planners and policymakers to mitigate flooding risks and improve resilience in vulnerable regions.

1. Introduction

Floods are among the most devastating natural disasters globally, causing significant human and economic losses annually [1,2,3,4,5,6,7]. Urban pluvial flooding (UPF), characterized by surface water accumulation due to intense rainfall overwhelming drainage systems, has become increasingly prevalent. This escalation is attributed to climate change-induced rainfall intensification and rapid urbanization [3,8,9,10,11,12,13,14,15,16,17,18,19]. The consequences of UPF include infrastructure damage, disruption of urban life, and substantial economic losses [20]. These impacts are projected to worsen with continued urban expansion and the increasing frequency of extreme weather events [21], underscoring the urgent need for innovative solutions to mitigate UPF risks.

In many developing countries, such as Pakistan, the assessment of flood risk is hindered by the lack of comprehensive drainage network data and historical flood records. Traditional flood mapping techniques, which often rely on statistical analyses and Geographic Information Systems (GIS), struggle to capture the dynamic nature of urban flooding, particularly in areas with outdated or inadequate drainage infrastructure [22]. These methods are heavily dependent on historical data and often fail to provide accurate assessments in data-scarce environments.

Recent literature indicates a predominant focus on traditional models for urban flooding assessment, including histogram-based, threshold, statistical, and GIS models [14,16,17,18,19]. While deep learning (DL) approaches have shown promise in various flood-related applications, their implementation in UPF assessment remains limited. Ref. [23] reviewed approximately 1038 research articles related to flood disasters, finding that only about 46 employed deep learning-based modeling for flood assessment. Moreover, practical implementations of DL models in UPF contexts are scarce [24,25,26,27,28].

The challenges posed by data scarcity and the dynamic nature of urban systems necessitate the development of new methods to improve flood susceptibility assessment and enhance urban resilience. Conventional approaches, such as Multi-Criteria Decision Analysis (MCDA) and GIS-based models, are widely used but often lack the adaptability required for modern urban environments [29,30,31,32,33,34,35]. There is a pressing need for more adaptive, data-driven solutions that can provide accurate assessments even in the absence of extensive historical data.

This study aims to develop an improved flood risk assessment framework tailored for data-scarce urban regions. By leveraging alternative data sources and advanced computational approaches, this research seeks to address the limitations of traditional flood modeling methods. The findings are expected to offer valuable insights for urban planners and policymakers, facilitating the implementation of more effective flood prevention and mitigation strategies. Additionally, this research underscores the importance of integrating modern data-driven methodologies to enhance flood resilience in vulnerable urban environments.

2. Dataset

This study integrates multiple datasets to support both traditional and advanced modeling approaches to address the challenges of urban pluvial flooding (UPF) assessment in data-scarce regions.

For traditional modeling, the high-resolution ALOS PALSAR Digital Elevation Model (DEM) with a 12.5 m spatial resolution was utilized. Compared to alternatives like ASTER GDEM (30 m) and SRTM (30 m), ALOS PALSAR offers superior spatial and vertical accuracy, with a vertical accuracy of approximately ±5 m. This higher resolution ensures more precise extraction of drainage networks, which is crucial for accurate flood modeling. Additionally, ALOS PALSAR is freely available, making it a valuable resource for regions with limited access to high-quality geospatial data.

Table 1. Summary of datasets used.

Dataset	Source/Resolution	Purpose	Relevance to Study	URL
ALOS PALSAR DEM	12.5 m resolution (Free)	Drainage network extraction	High-resolution terrain modeling	https://www.eorc.jaxa.jp/ALOS/en/dataset/alos_open_and_free_e.htm, accessed on 7 May 2025
ASTER GDEM	30 m resolution (Comparison)	Elevation data (comparison)	Lower resolution; less accurate	https://asterweb.jpl.nasa.gov/gdem.asp, accessed on 7 May 2025
SRTM DEM	90 m resolution (Comparison)	Elevation data (comparison)	Coarser resolution; less precise	https://www.earthdata.nasa.gov/data/instruments/srtm, accessed on 7 May 2025
Sentinel-2 Imagery	10 m resolution (Free)	Land use/land cover classification	Key for understanding urbanization	https://www.esa.int/Applications/Observing_the_Earth/Copernicus/Sentinel-2, accessed on 7 May 2025
GISAI and GHSL Data	Varies	Impervious surface area mapping	Important for runoff analysis	https://data.jrc.ec.europa.eu/collection/ghsl, accessed on 7 May 2025
Precipitation Data	Government sources (Monthly)	Rainfall intensity and variability	Identifies rainfall patterns	https://nwfc.pmd.gov.pk/new/rainfall.php, accessed on 7 May 2025
Slope and Aspect	Derived from DEM	Terrain analysis	Determines water flow direction	Derived
Historical UPF Events	2–3 past events (Synthetic data included)	DRNN model training	Captures temporal flood patterns	https://www.ndma.gov.pk/urbanflooding/sdi, accessed on 7 May 2025

provide the information of dataset used particularly, for the implementation of Multi-Criteria Decision Analysis (MCDA) using the Analytical Hierarchy Process (AHP), multiple datasets were integrated, including impervious surface area (ISA), land use/land cover (LULC), rainfall, slope, aspect, and other relevant factors. Each dataset contributes unique insights into the hydrological and urban characteristics of the study area, facilitating a comprehensive evaluation of flood susceptibility. Figure 1 provides an overview of the datasets used.

These datasets collectively form the foundation for both traditional and deep learning-based flood susceptibility modeling. The use of high-resolution DEM data ensures precise drainage network extraction, while artificial data generated through GANs helps mitigate the challenges posed by limited historical flood data. This integrated approach ensures that the models developed are robust, accurate, and capable of addressing the complexities of urban pluvial flooding in data-scarce regions.

3. Methodology

This study integrates deep learning (DL) models with traditional approaches to address the challenges of urban pluvial flooding (UPF) mapping in data-scarce regions. Lahore, Pakistan, a rapidly urbanizing city frequently impacted by monsoonal UPF, was selected as the study area. The city’s inadequate drainage infrastructure and extensive impervious surfaces exacerbate its vulnerability to flooding.

The research workflow for urban pluvial flooding (UPF) mapping integrates multiple datasets and methodologies to address challenges in data-scarce regions. Figure 2 provides a detailed workflow diagram, illustrating how the modules interact and the sequence of operations.

3.1. Data Collection and Pre-Processing

Key datasets included the high-resolution ALOS PALSAR Digital Elevation Model (DEM) with a 12.5 m spatial resolution, utilized for natural drainage network extraction. Sentinel-2 imagery facilitated land use/land cover (LULC) classification, while GISAI and GHSL datasets provided information on impervious surface mapping. Additionally, precipitation data spanning from 2001 to 2022 were obtained from government sources for flood analysis. These datasets served as foundational inputs for both traditional and deep learning-based modeling approaches.

The ALOS PALSAR DEM data were processed using ArcGIS Pro 3.0 to extract natural drainage networks. Hydrological tools generated flow direction and accumulation maps, identifying potential flooding hotspots. These outputs also served as validation data for the deep learning models.

3.2. Synthetic Data Generation with GANs

To address data scarcity, Generative Adversarial Networks (GANs) were employed to generate synthetic drainage network data (Figure 3). GANs consist of two neural networks: a generator that creates synthetic data and a discriminator that evaluates the authenticity of the generated data. This adversarial training process enables the generation of realistic synthetic datasets. The synthetic drainage data produced by GANs supplemented the natural drainage data, ensuring robust inputs for the deep learning models.

3.3. Deep Learning Model (DRNN)

The Deep Recurrent Neural Network (DRNN) was utilized to model temporal dependencies using historical precipitation data and synthetic drainage data from GANs. The DRNN architecture comprised multiple stacked RNN layers and incorporated techniques such as layer normalization and Leaky Rectified Linear Unit (Leaky ReLU) activation functions to enhance prediction accuracy for flood-prone areas.

3.4. Multi-Criteria Decision Analysis (MCDA)

In parallel, the Analytical Hierarchy Process (AHP), a traditional Multi-Criteria Decision Analysis (MCDA) technique, was applied. This approach integrated multiple factors, including rainfall intensity, drainage density, and impervious surface area, to assess flood susceptibility zones. The AHP involved constructing a pairwise comparison matrix to assign weights to each criterion based on their relative importance. The MCDA results served as a benchmark for comparison with the deep learning outputs.

3.5. Hyper-Parameter Optimization

Hyper-parameter optimization was conducted for both the DRNN and GAN models to enhance predictive accuracy and model stability. For the DRNN, a grid search was performed over the following ranges: number of hidden layers (1–3), neurons per layer (50–200), learning rate (0.001–0.01), and dropout rate (0.1–0.5). The optimal configuration was determined based on validation loss and predictive performance. For the GAN, training epochs (100–500), batch size (32–128), and learning rates (0.0001–0.001) for both the generator and discriminator were varied. Final settings were selected based on the quality of generated data, assessed using Structural Similarity Index Measure (SSIM) and Root Mean Square Error (RMSE) metrics. This systematic tuning process ensured the robustness and reliability of the models.

4. Results

4.1. Impervious Surface Area (ISA) and Land Use/Land Cover (LULC) Analysis

Urbanization significantly influences urban pluvial flooding (UPF) by altering natural drainage systems and increasing impervious surface areas. Sentinel-2 imagery was used to classify land use and land cover (LULC), identifying key categories such as urban, vegetation, and water bodies. The results indicate that rapid urban expansion has led to a significant increase in impervious surfaces, reducing the natural infiltration capacity and exacerbating surface runoff. The integration of GISAI and GHSL datasets provided an accurate delineation of impervious surface areas (ISA), revealing that high-density urban zones correspond with the most flood-prone regions. Figure 4 illustrates the classified LULC map, while Figure 4 highlights impervious surface distribution, showing the temporal variation for last three decades.

The preliminary results of this study provide key insights into impervious surface area (ISA), land use/land cover (LULC), precipitation patterns, and drainage network derivation, each contributing to a comprehensive understanding of urban pluvial flooding (UPF). The impervious surface area was delineated by combining GISAI and GHSL datasets, using an intersection-based approach to yield an accurate representation of impervious surfaces in the study area. Figure 5, shows the land use map 2022 with four majorland use types.

4.2. Drainage Network Extraction and Flood Hotspots Identification

The natural drainage network was extracted using ALOS PALSAR DEM data, processed through hydrological tools in ArcGIS Pro. Flow direction and accumulation analysis were performed to delineate natural water movement and identify potential drainage pathways. The results indicate that low-lying areas with poor drainage infrastructure are more prone to flooding, particularly in rapidly urbanizing districts. Figure 6 highlights flood hotspots where the drainage system is insufficient to accommodate heavy precipitation events. These findings emphasize the urgent need for improved drainage infrastructure in high-risk urban zones highlighted by red color in map.

Rainfall data spanning from 2001 to 2022 were normalized for each month and analyzed (Figure 7), revealing an increase in precipitation intensity during monsoon seasons. These findings underscore the heightened vulnerability of urban areas to pluvial flooding due to climatic changes, which align with trends observed in other developing regions.

4.3. Synthetic Data Generation with GANs and Data Quality Assessment

Due to the scarcity of high-resolution historical flood data, Generative Adversarial Networks (GANs) were implemented to generate synthetic drainage networks. The GAN model was trained using real drainage data, enabling it to produce synthetic datasets that closely mimic natural hydrological patterns.

The GAN was trained using real drainage networks extracted from ALOS PALSAR DEM data for multiple urban sectors within Lahore. The architecture followed a Deep Convolutional GAN (DCGAN) structure. The generator consisted of transposed convolutional layers with batch normalization and ReLU activations, while the discriminator was built with convolutional layers using Leaky ReLU and dropout for regularization. Binary cross-entropy was used as the loss function for both networks, optimized using the Adam optimizer (learning rate = 0.0002, β₁ = 0.5). The model was trained for 200 epochs with a batch size of 64 and image input size of 128 × 128 pixels. These configurations were selected based on standard best practices and preliminary tuning to ensure model stability and convergence.

To evaluate the quality of the generated data, a comparison was made between real and synthetic drainage networks using the structural similarity index measure (SSIM) and root mean square error (RMSE). The results indicate that the GAN-generated data achieved an SSIM score of 0.92 and an RMSE of 0.14, demonstrating high similarity to real-world data.

Figure 8 provides a visual comparison of real and synthetic drainage networks, while

Table 2. Validation metrics of GAN-generated data.

Metric	Value	Description
SSIM (Structural Similarity Index Measure)	0.92	Measures similarity between real and synthetic data (closer to 1 indicates high similarity).
RMSE (Root Mean Square Error)	0.14	Indicates average deviation of synthetic data from real data (lower is better).
MAE (Mean Absolute Error)	0.11	Measures average absolute difference between real and synthetic data.
Data Coverage (%)	95%	Percentage of drainage features accurately captured in synthetic data.

presents validation metrics confirming the reliability of GAN-generated data. These results indicate that GANs can effectively supplement missing drainage information in data-scarce regions, enhancing the accuracy of UPF modeling.

In addition to SSIM and RMSE, physical consistency checks were performed. The drainage network characteristics, including flow direction and accumulation patterns, were compared with those derived independently from high-resolution DEM data. These checks confirmed that the GAN-generated drainage networks closely match the physical attributes expected in natural drainage systems. Furthermore, comparative experiments were conducted by validating the synthetic drainage networks against manually extracted networks from an independent dataset, demonstrating comparable performance across key metrics.

4.4. Deep Learning Model (DRNN) Results for Flood Prediction

To predict flood-prone areas, a Deep Recurrent Neural Network (DRNN) was trained using precipitation data, LULC, and drainage network information. The model successfully identified high-risk flooding zones by capturing spatiotemporal dependencies in rainfall intensity and land characteristics. Model evaluation metrics showed an overall precision of 85%, a recall of 83%, and an F1-score of 84%, indicating a robust predictive performance. The inclusion of GAN-generated synthetic data improved the model’s accuracy by 7%, compared to models trained solely on real data. Figure 8 illustrates the drainage network generated by DRNN, highlighting vulnerable urban areas, while

Table 3. (a) DRNN performance metrics. (b) DRNN regression metrics.

(a)
Metric	Real Data Only	Real + Synthetic Data	Improvement (%)
Precision	78%	85%	+7%
Recall	76%	83%	+7%
F1-Score	75%	84%	+9%
Accuracy	77%	86%	+9%
(b)
Metric		Value
MSE		0.087
RMSE		0.295
R2		0.81

(a) summarizes the model’s performance metrics.

These results confirm the effectiveness of deep learning approaches in UPF prediction, particularly when combined with synthetic data.

In addition to classification metrics, we evaluated the DRNN model using regression-based performance indicators to better reflect continuous flood susceptibility prediction. The model achieved a Mean Squared Error (MSE) of 0.087, a Root Mean Squared Error (RMSE) of 0.295, and an R² score of 0.81, indicating a strong fit between predicted and reference data (Table 3 (b)). These results confirm the model’s capability to generalize well across different spatial conditions and flooding intensities.

4.5. Comparison of DRNN and Multi-Criteria Decision Analysis (MCDA) Results

To evaluate the effectiveness of the DRNN model, its results were compared with those from the traditional Multi-Criteria Decision Analysis (MCDA) approach. While both models successfully identified major flood-prone zones, the DRNN model exhibited higher spatial precision and adaptability to varying precipitation patterns. The comparison revealed that DRNN outperformed MCDA in capturing localized flood susceptibility, particularly in areas where historical flood data were scarce.

Table 4. DRNN vs. MCDA performance metrics.

Metric	DRNN	MCDA	Difference
Precision	85%	78%	+7%
Recall	83%	75%	+8%
F1-Score	84%	76%	+8%
Area Under Curve (AUC)	0.88	0.81	+0.07

presents a side-by-side comparison of DRNN and MCDA and quantifies their respective performance metrics. Expert validation further confirmed that DRNN predictions aligned more closely with observed flood events. These findings highlight the advantages of deep learning in dynamic flood mapping, particularly for regions with complex urban hydrology.

4.6. Uncertainties, Errors, and Model Limitations

While the results demonstrate the potential of deep learning for UPF mapping, several uncertainties and limitations must be acknowledged. The accuracy of the models is dependent on the quality and resolution of input datasets, such as DEMs, LULC classifications, and precipitation records. Additionally, while GAN-generated synthetic data improved model performance, biases may arise if the training dataset does not adequately capture the full variability of real-world drainage systems. Overfitting in DRNN models is another concern, particularly when trained on limited historical flood data. Furthermore, the computational cost of training deep learning models may be a barrier for real-time flood prediction applications in resource-constrained environments.

To mitigate these limitations, future research should focus on incorporating higher-resolution datasets, expanding GAN training with diverse urban flood scenarios, and testing model generalizability in different geographic regions.

Table 5. Summary of model uncertainties and errors.

Source of Uncertainty/Error	Impact on Results	Mitigation Strategies
Input Data Resolution (DEM, LULC)	Reduced accuracy in small-scale flood mapping	Use higher-resolution datasets (e.g., LiDAR)
GAN Data Biases	Potential misclassification due to synthetic data inaccuracy	Expand GAN training dataset diversity
DRNN Overfitting	Reduced model generalization to new areas	Implement regularization and cross-validation
Precipitation Data Aggregation	Loss of short-term rainfall variability	Integrate real-time rainfall data
Computational Cost	Limited application in resource-constrained settings	Develop optimized, lightweight model versions

summarizes the key sources of uncertainty in this study and their potential impact on model predictions.

A key limitation of this study is the exclusive focus on natural drainage networks derived from DEM data. Due to the unavailability of detailed maps and data on artificial drainage systems (e.g., stormwater infrastructure), these features were not incorporated into the model. However, surface permeability was partially addressed using impervious surface maps from GISAI and GHSL datasets, which capture urbanization patterns. Additionally, long-term normalized precipitation data were used to reflect seasonal and interannual rainfall variability. Future research should aim to integrate artificial drainage infrastructure and simulate dynamic hydrological processes using rainfall-runoff models to improve realism in urban flood modeling.

A notable limitation of this study is the reliance on remote sensing and simulation data, necessitated by the unavailability of in situ water level measurements and detailed historical flood records in the study area. The integration of such observed data is crucial for improving model calibration and validation. Future research should aim to incorporate actual hydrological measurements to enhance the model’s accuracy and reliability in flood risk assessment.

These insights will be crucial in refining deep learning methodologies for improved flood risk assessment and urban resilience planning.

5. Discussion

This study demonstrates the efficacy of integrating deep learning (DL) models with traditional methods for urban pluvial flooding (UPF) assessment in data-scarce regions. The application of Deep Convolutional Generative Adversarial Networks (DCGANs) to simulate artificial drainage networks effectively addresses data scarcity challenges. When combined with Digital Elevation Model (DEM)-based natural drainage mapping, this hybrid approach offers a comprehensive framework for developing robust and adaptive flood mitigation strategies.

Comparative analyses between DL models, particularly Deep Recurrent Neural Networks (DRNNs), and traditional Multi-Criteria Decision Analysis (MCDA) techniques, such as the Analytical Hierarchy Process (AHP), indicate the superior accuracy and robustness of DRNNs. These findings align with previous research highlighting the advantages of DL methods in managing complex environmental and urban systems [23,36].

Urbanization and climatic changes are critical factors exacerbating UPF risks. The study underscores how increased impervious surfaces and outdated drainage systems heighten flooding vulnerabilities, necessitating innovative, data-driven approaches. Additionally, intensified monsoon precipitation due to climatic shifts further complicates flood management, calling for adaptive solutions.

Despite promising outcomes, data scarcity remains a significant hurdle, particularly for training DL models [24,25,26,27,28]. The generation of synthetic datasets using GANs and their integration into MCDA frameworks present a viable path forward. Future work will focus on validating these models with real-world flood event data and extending the methodology to other data-scarce regions globally.

Implementing DL models for UPF is challenging due to complex interactions within natural and built environments, such as drainage systems and urban infrastructure. Data limitations in the study area encompass both hydraulic datasets and historical UPF event records. This study addresses these gaps by developing synthetic data to train DL models, enhancing their predictive capabilities.

The integration of GANs and DRNNs enhances UPF prediction in data-scarce regions. GANs effectively generate synthetic drainage network data, filling critical data gaps and enabling more accurate flood susceptibility modeling. These improved predictions can inform urban planning decisions and the development of early warning systems, contributing to more effective flood mitigation strategies.

However, several uncertainties affect the reliability of model predictions. The accuracy of results depends on the quality and resolution of input datasets, such as ALOS PALSAR DEM and Sentinel-2 imagery. Limitations inherent in these datasets, like resolution constraints and potential misclassification in land use/land cover (LULC) data, can introduce errors into both GAN-generated synthetic data and DRNN flood predictions. Moreover, synthetic data may introduce biases if training data do not fully capture the variability of natural drainage patterns.

Potential sources of error include: (1) limitations in input data, such as DEM spatial resolution and LULC classification accuracy, affecting the fidelity of extracted drainage networks; (2) GAN-generated data may not perfectly replicate real-world drainage system diversity, leading to model bias; (3) limited historical flood events for training increase the risk of overfitting in the DRNN model, reducing generalizability; and (4) temporal aggregation of precipitation data into broader intervals may obscure short-term rainfall extremes critical for flash flood prediction.

To mitigate these uncertainties, future research should incorporate higher-resolution datasets, such as LiDAR-based DEMs, and expand the diversity of GAN training datasets. Integrating real-time data could enhance the dynamic performance of DRNN models, refining predictive capabilities and reducing uncertainties. These improvements are crucial for ensuring the models’ applicability across diverse urban environments with varying hydrological and urbanization characteristics.

The proposed framework, integrating GANs with DRNNs, offers a promising solution for UPF prediction in data-scarce regions. By generating synthetic drainage network data, it enhances flood risk assessment, benefiting rapidly urbanizing cities with inadequate drainage infrastructure. The model’s adaptability allows integration into urban planning and early warning systems.

However, the framework’s accuracy depends on input data quality, such as DEMs and LULC classifications, which may introduce errors. While GAN-generated data address data scarcity, they may also introduce biases if training data lack variability [37]. Additionally, the high computational cost of DRNNs may limit applicability in resource-constrained settings, and generalization to different urban environments may require recalibration. Future research should focus on improving data resolution, expanding GAN training diversity, and optimizing model efficiency. Testing the framework in diverse urban contexts will enhance its robustness and practical applicability.

6. Conclusions

Urban pluvial flooding (UPF) remains a pressing challenge, particularly in regions where data scarcity hampers effective risk assessment and management. This study introduces an integrative framework that combines traditional methodologies with advanced deep learning (DL) techniques to address these challenges. By leveraging high-resolution ALOS PALSAR DEM data for drainage network extraction and employing Generative Adversarial Networks (GANs) for synthetic data generation, the approach effectively mitigates data limitations. Furthermore, the incorporation of Deep Recurrent Neural Networks (DRNNs) enhances flood susceptibility mapping by capturing temporal patterns in precipitation and urban features.

The findings indicate that DL models, particularly DRNNs augmented with GAN-generated datasets, surpass traditional Multi-Criteria Decision Analysis (MCDA) methods in both accuracy and robustness. This underscores the compounded impacts of urbanization and climate change on UPF risks and highlights the necessity for innovative, data-driven solutions.

Beyond technical advancements, the study offers actionable insights for urban planners. Integrating model outputs into planning strategies—such as targeted infrastructure improvements and land-use regulations—can significantly reduce UPF risks and bolster long-term resilience in vulnerable areas.

Future research should focus on validating the proposed models with real-world flood events and assessing their applicability across diverse urban environments. Expanding datasets to include real-time monitoring data and enhancing model generalizability across varying hydrological settings will further strengthen the framework’s reliability. By providing a scalable, data-driven solution for proactive flood management, this framework contributes to the development of more resilient and sustainable urban landscapes.