Karapetyan et al. (2026) Deep vision-based framework for coastal flood prediction under sea level rise and shoreline protection

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-01-02
• Authors: A.P. Karapetyan, Aaron C. Chow, Samer Madanat
• DOI: 10.1038/s41598-025-33803-z

Research Groups

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

Short Summary

The study develops a vision-based deep learning framework, featuring a lightweight CNN architecture named CASPIAN, to predict high-resolution coastal flood depths under sea level rise and various shoreline protection scenarios. The framework achieves accuracy comparable to physics-based hydrodynamic models while providing a $10^8$ times increase in inference speed.

Objective

• To develop a performant and computationally efficient deep learning surrogate model capable of predicting peak coastal floodwater levels in low-data settings, specifically accounting for sea level rise (SLR) and the hydrodynamic impacts of shoreline fortifications (seawalls).

Study Configuration

• Spatial Scale: Coastal Abu Dhabi, UAE; 12,066 nearshore locations with horizontal resolutions up to 30 m in urban areas, mapped to a $1024 \times 1024$ mesh grid.
• Temporal Scale: 0.5 m Sea Level Rise (SLR) projection (representative of 2050 estimates); training data based on 3-month hydrodynamic simulation periods including Shamal wind events.

Methodology and Data

• Models used:
  • Physics-based (Ground Truth): A coupled model consisting of Delft3D (tides/hydrodynamics), SWAN (spectral waves), and an empirical run-up model.
  • Deep Learning Surrogates: CASPIAN (Cascaded Pooling and Aggregation Network), SWIN-Unet (Transformer-based), and Attention U-Net.
  • Benchmarks: Kriging with Principal Component Analysis (PCA), Support Vector Regression (SVR), and Lasso with polynomial features.
• Data sources: Synthetic flood depth maps generated via the coupled physics-based model; ERA5 database for wind and atmospheric pressure forcing; TPXO8 Ocean Atlas for tidal gauge calibration.

Main Results

• Accuracy: CASPIAN achieved high precision, with approximately 97% of predicted floodwater levels having an absolute error of $\le 0.1$ m.
• Efficiency: The model reduced the computational time for generating flood maps from approximately 72 hours (physics-based) to milliseconds ($10^8$ speedup).
• Model Compactness: CASPIAN is a lightweight architecture with only 0.36 million parameters, making it trainable on a single GPU.
• Generalization: The vision-based approach, supported by the Cutout augmentation technique, allowed the model to perform effectively even with a limited training set of approximately 100 samples.

Contributions

• Framework Design: Recasts high-dimensional coastal flood regression as a computer vision image-to-image translation task, enabling the use of advanced vision architectures and augmentation.
• Novel Architecture: Introduction of CASPIAN, a CNN specifically stylized for climate-adaptation-aware flood prediction using cascaded pooling and aggregated residual transformations.
• Open Data: Provision of a first-of-its-kind benchmark dataset containing 174 high-resolution synthetic flood depth maps for the coast of Abu Dhabi under 0.5 m SLR.
• Open Source: Public release of the complete source code and trained models to facilitate coastal engineering research.

Funding

• Research conducted at New York University Abu Dhabi. (Specific project reference codes were not explicitly provided in the source text).

Citation

@article{Karapetyan2026Deep,
  author = {Karapetyan, A.P. and Chow, Aaron C. and Madanat, Samer},
  title = {Deep vision-based framework for coastal flood prediction under sea level rise and shoreline protection},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-025-33803-z},
  url = {https://doi.org/10.1038/s41598-025-33803-z}
}

Generated by BiblioAssistant using gemini-3-flash-preview (Google API)