Hassan et al. (2026) Climate adaptation-aware flood prediction for coastal cities using Deep Learning

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-03-11
• Authors: Bilal Hassan, Areg Karapetyan, Aaron C. Chow, Samer Madanat
• DOI: 10.5194/hess-30-1333-2026

Research Groups

• Division of Engineering, New York University Abu Dhabi, Abu Dhabi, United Arab Emirates

Short Summary

This study develops a novel, lightweight Convolutional Neural Network (CNN)-based model, CASPIAN-v2, for rapid and accurate prediction of high-resolution coastal flooding in urban areas under various sea-level rise scenarios and shoreline adaptation strategies. The model significantly outperforms state-of-the-art methods, reducing mean absolute error by nearly 20%, and offers a scalable tool for climate adaptation planning.

Objective

• To develop a novel, lightweight Convolutional Neural Network (CNN)-based model (CASPIAN-v2) for accurate and rapid prediction of high-resolution coastal flooding under various sea-level rise (SLR) scenarios and shoreline adaptation strategies.
• To demonstrate the model's ability to generalize across diverse geographical contexts and multiple SLR scenarios.
• To provide a scalable and practical tool for coastal flood management and strategic planning, enabling the exploration of optimal spatial configurations for flood defenses.

Study Configuration

• Spatial Scale: Two urban coastal regions: Abu Dhabi (United Arab Emirates) with 17 operational landscape units (OLUs), and San Francisco Bay Area (United States) with 30 OLUs. Flood prediction is performed on a 1024x1024 spatial grid, with underlying hydrodynamic models using a 30 meter grid resolution.
• Temporal Scale: Future extreme flooding scenarios under various sea-level rise (SLR) projections (0.5 meter, 1.0 meter, and 1.5 meter). Hydrodynamic simulations for training data covered 3-month periods, capturing hourly inland inundation. SLR projections are consistent with mid-century (2050-2100) IPCC AR6 scenarios.

Methodology and Data

• Models used:
  • Deep Learning: CASPIAN-v2 (novel lightweight Convolutional Neural Network - CNN), CASPIAN (previous version), Compact Convolutional Transformers (CCT), Attention-Unet, Swin-Unet, Multi-layer Perceptron (MLP).
  • Traditional Machine Learning: Naïve model, Random Forest, Linear Regression, Extreme Gradient Boosting, Support Vector Regression, Lasso Regression with Polynomial Features, Kriging with Principal Component Analysis.
  • Physics-based Hydrodynamic Simulators (for ground truth data generation): Delft3D, SWAN (for wind and wave effects in Abu Dhabi).
• Data sources:
  • Ground truth flood data generated from high-fidelity physics-based hydrodynamic simulations (Delft3D, SWAN).
  • High-resolution bathymetry and digital elevation models (DEM) from TanDEM-X, Landsat-8, and Nautical Charts.
  • ERA5 meteorological data for forcing hydrodynamic models in Abu Dhabi.
  • Validation of hydrodynamic models against real-world observations from tidal gauges (9 locations in San Francisco Bay, 196 locations in the Arabian Gulf).
  • Two new comprehensive datasets from Abu Dhabi (0.5 meter SLR) and San Francisco (0.5, 1.0, 1.5 meter SLR) covering various shoreline adaptation scenarios.
  • Data augmentation techniques (random spatial cutouts and scaling factors) applied to expand training datasets.

Main Results

• The proposed CASPIAN-v2 model significantly outperforms state-of-the-art (SOTA) methods, reducing the Mean Absolute Error (AMAE) in predicted flood depth maps by 19.96% compared to the second-best deep learning model (CASPIAN) and 51.65% compared to the best traditional machine learning model (Lasso with polynomial features).
• CASPIAN-v2 achieves high spatial accuracy, with a Dice Similarity Coefficient (DSC) of 0.8437, representing a 31.05% improvement over Lasso and 2.13% over the original CASPIAN model.
• The model demonstrates high non-inundated prediction accuracy (Acc[0] of 99.39%) and a very low false negative rate (0.81%), indicating reliable identification of both dry and flooded areas despite data imbalance.
• Computational efficiency is dramatically improved: CASPIAN-v2 can generate predictions for 72 test scenarios in less than 16 seconds on a single GPU, compared to approximately 2763 hours (115 days) for physics-based simulations.
• The model is lightweight with 0.38 million parameters, making it highly scalable.
• CASPIAN-v2 exhibits strong generalizability across different sea-level rise scenarios (0.5 meter and 1.5 meter for San Francisco) and complex, unseen shoreline adaptation configurations with minimal fine-tuning.
• Explainable AI (Grad-CAM) confirms the model's focus on vulnerable, unprotected shoreline segments, enhancing interpretability.
• Predictive uncertainty maps, derived from a deep ensemble, show a strong spatial correlation between high uncertainty and higher absolute prediction errors, providing a valuable tool for identifying less reliable predictions (e.g., uncertainty > 0.75 indicates potential error > 0.5 meter).

Contributions

• Development of CASPIAN-v2, a novel, lightweight Convolutional Neural Network (CNN) architecture for high-resolution coastal flood prediction that integrates sea-level rise (SLR) data and generalizes across diverse geographical regions and adaptation scenarios.
• Significant reduction in computational time for flood prediction, transforming a months-long simulation effort into a near-instantaneous task, enabling large-scale scenario exploration for coastal resilience planning.
• Creation and release of two new, comprehensive datasets for Abu Dhabi and San Francisco, covering multiple SLR scenarios and shoreline adaptation strategies, to foster further research.
• Rigorous benchmarking of CASPIAN-v2 against a wide array of state-of-the-art machine learning and deep learning models, demonstrating superior accuracy and generalization capabilities.
• Integration of Explainable AI (Grad-CAM) and deep ensemble methods to provide interpretability and quantify predictive uncertainty, enhancing trust and supporting risk-informed decision-making for urban policymakers.
• Open-sourcing of the code and datasets to promote reproducibility and further research in climate adaptation-aware flood prediction.

Funding

• New York University Abu Dhabi (NYUAD)

Citation

@article{Hassan2026Climate,
  author = {Hassan, Bilal and Karapetyan, Areg and Chow, Aaron C. and Madanat, Samer},
  title = {Climate adaptation-aware flood prediction for coastal cities using Deep Learning},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-1333-2026},
  url = {https://doi.org/10.5194/hess-30-1333-2026}
}