Chakrabortty et al. (2025) Urban Flood Susceptibility Assessment in Arid Environment Using a Novel Hybrid Deep Learning Approach

Identification

• Journal: Earth Systems and Environment
• Year: 2025
• Date: 2025-11-21
• Authors: Rabin Chakrabortty, Tarig Ali, Mohamed Yehia Abouleish, Serter Atabay, Norita Ahmad, Raafat Aburukba, Gowhar Meraj, J. Nave, Shrouq Maher Al-Etoom
• DOI: 10.1007/s41748-025-00933-3

Research Groups

• Department of Civil Engineering, College of Engineering, American University of Sharjah, United Arab Emirates
• Energy, Water, and Sustainable Environment Research Center, College of Engineering, American University of Sharjah, United Arab Emirates
• Department of Biology, Chemistry and Environmental Sciences, College of Arts and Sciences, American University of Sharjah, United Arab Emirates
• School of Business Administration, American University of Sharjah, United Arab Emirates
• Department of Computer Science and Engineering, American University of Sharjah, United Arab Emirates
• Department of Built Environment, College of Science and Technology, North Carolina A&T State University, Greensboro, USA
• Department of Industrial Engineering, College of Engineering, American University of Sharjah, United Arab Emirates

Short Summary

This study develops Hydro-TransformerNet, a novel hybrid deep learning framework for urban flood susceptibility mapping in data-scarce arid environments, demonstrating strong predictive performance (AUC of 0.945) by integrating spatial, temporal, and hydrologically guided attention mechanisms. The model effectively identifies flood-prone areas in Sharjah, UAE, using remote sensing data and synthetic flood masks, providing a scalable and interpretable tool for urban planning and disaster mitigation.

Objective

• To develop Hydro-TransformerNet, a data-efficient and transferable hybrid deep learning framework for urban flood susceptibility assessment in arid and semi-arid environments, specifically applied to the Emirate of Sharjah, UAE, under data-scarce conditions.
• To combine spatio-temporal learning, operate in the absence of historical flood masks, and provide spatial explainability through hydrologically guided attention.

Study Configuration

• Spatial Scale: Emirate of Sharjah, United Arab Emirates, covering diverse landforms from coastal plains to mountainous catchments. Input data processed to a uniform spatial resolution of 10 meters.
• Temporal Scale: Focuses on urban flash flood susceptibility, which are low-frequency but high-impact events. The model simulates temporal dynamics using synthetic sequences to mimic the persistence of hydrometeorological conditions over time, addressing the lack of real-time series data.

Methodology and Data

• Models used:
  • Hydro-TransformerNet: A novel hybrid deep learning architecture integrating a ResNet34-based convolutional encoder for fine-scale spatial feature extraction, a temporal transformer module for simulating sequential hydrological processes, and a hydrologically guided attention mechanism.
  • Benchmarked against: Traditional U-Net and Attention-augmented U-Net (U-Net + Attention).
  • Feature importance analysis: SHapley Additive exPlanations (SHAP) and Random Forest (RF).
• Data sources:
  • Digital Elevation Model (DEM): ALOS PALSAR (10 m spatial resolution).
  • Slope: Derived from DEM.
  • Topographic Wetness Index (TWI): Derived from DEM.
  • Normalized Difference Vegetation Index (NDVI): Sentinel-2 satellite imagery.
  • Normalized Difference Water Index (NDWI): Sentinel-2 satellite imagery.
  • Land Use/Land Cover (LULC): Supervised classification of Sentinel-2 imagery.
  • Impervious Surface: Derived from spectral mixture analysis of Sentinel-2 imagery.
  • Soil Type: FAO Harmonized World Soil Database.
  • Rainfall: Local meteorological records, spatially interpolated using Inverse Distance Weighting (IDW).
  • Synthetic binary flood masks: Developed using expert knowledge of hydrological patterns, low-lying urban areas, and rule-based combinations of hydrological and topographical parameters due to the absence of historical flood information.

Main Results

• The Hydro-TransformerNet model achieved strong predictive performance with an Area Under the Curve (AUC) of 0.945.
• Quantitative evaluation metrics showed an overall Accuracy of 0.916, Sensitivity of 0.914, Specificity of 0.918, F1-Score of 0.915, and a Kappa coefficient of 0.832.
• Training loss consistently decreased from approximately 0.9 to 0.3 over five epochs, indicating stable optimization. Training accuracy improved from 93% to over 99%, with validation accuracy remaining consistent between 98.0% and 99.2%, suggesting minimal overfitting.
• SHAP-based feature importance analysis revealed Elevation (0.450), Slope (0.197), and Rainfall (0.153) as the most dominant factors influencing flood susceptibility. Features like NDVI, NDWI, and Soil Type showed relatively low importance.
• The predicted flood susceptibility patterns align well with known flood hotspots, hydrological theory, and municipal reports, concentrating very high susceptibility in low-lying coastal/sub-coastal zones, wadis, and floodplains at the foothills of the Hajar Mountains.
• The model's calibration curve demonstrated good alignment between predicted probabilities and actual outcomes.

Contributions

• Introduces Hydro-TransformerNet, a novel hybrid deep learning framework that synergistically integrates CNNs for spatial feature extraction, transformer-based encoders for temporal pattern learning, and a hydro-physically guided attention mechanism.
• Addresses the critical need for high-resolution, transferable, and explainable flood susceptibility models capable of functioning in data-scarce arid and semi-arid regions without historical flood records, using satellite-derived environmental variables.
• Develops and validates a methodology for generating synthetic binary flood masks as a proxy for supervised learning in the absence of ground truth flood data.
• Provides spatial explainability through a hydrologically guided attention module and SHAP analysis, enhancing interpretability for environmental science applications.
• Generates GIS-compatible flood susceptibility maps, enabling direct integration with urban planning and disaster mitigation platforms.
• Offers a scalable, data-efficient, and interpretable deep learning solution for early-warning systems, infrastructure planning, and climate resilience management in rapidly developing arid cities.

Funding

• American University of Sharjah, UAE, Project number: AUS (PDFA25).

Citation

@article{Chakrabortty2025Urban,
  author = {Chakrabortty, Rabin and Ali, Tarig and Abouleish, Mohamed Yehia and Atabay, Serter and Ahmad, Norita and Aburukba, Raafat and Meraj, Gowhar and Nave, J. and Al-Etoom, Shrouq Maher},
  title = {Urban Flood Susceptibility Assessment in Arid Environment Using a Novel Hybrid Deep Learning Approach},
  journal = {Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s41748-025-00933-3},
  url = {https://doi.org/10.1007/s41748-025-00933-3}
}