Blay et al. (2025) Geospatial and Deep Learning Approaches for Modeling Floodwater Depth in Urbanized Areas

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-12-24
• Authors: Jeffrey Blay, Leila Hashemi-Beni
• DOI: 10.3390/rs18010060

Research Groups

• Built Environment Department, College of Science and Technology, North Carolina A&T State University, Greensboro, NC, USA

Short Summary

This study developed a deep learning framework using geospatial and deep learning approaches to model floodwater depth in urbanized areas, finding that a lightweight ResNet18 architecture with terrain-derived predictors achieved high accuracy and spatial coherence, demonstrating potential for rapid flood assessment in data-scarce regions.

Objective

• To develop and evaluate deep learning-based models for urban floodwater depth estimation by leveraging hydrostatic principles and physically relevant variables derived from remote sensing sources.
• To create a practical regression framework integrating flood-inundated extent with Digital Terrain Models (DTMs) and hydrologically relevant topographic features (e.g., slope, curvature, Topographic Wetness Index) to produce interpretable and spatially coherent depth estimates in urban settings.

Study Configuration

• Spatial Scale: 12 urban and peri-urban areas in Southeastern United States (North Carolina and South Carolina), collectively spanning approximately 101.41 square kilometers. Data layers were resampled to a uniform 1 meter resolution and processed into 256 by 256 pixel patches. Model generalization was tested on an unseen catchment of approximately 2.16 square kilometers.
• Temporal Scale: Post-event conditions following Hurricane Matthew (October 10-15, 2016) and Hurricane Florence (September 18, 2018). Remote sensing imagery captured shortly after these events, with ground truth High Water Mark (HWM) data collected days later.

Methodology and Data

• Models used:
  • Convolutional Neural Networks (CNNs): ResNet-18, ResNet-34, ResNet-50
  • Transformer-based model: Swin UNet
  • Pre-trained U-Net (for flood extent delineation from aerial imagery)
• Data sources:
  • Post-flood aerial imagery (NOAA Storms Archive, approximately 0.25 meter spatial resolution).
  • Digital Terrain Models (DTM) derived from LiDAR point clouds (North Carolina Emergency Management Spatial Data Portal, USGS 3D Elevation Program (3DEP), approximately 1 meter spatial resolution).
  • High Water Mark (HWM) data (Federal Emergency Management Agency (FEMA) in partnership with USGS).
  • Derived terrain features: elevation, slope, curvature, and Topographic Wetness Index (TWI).

Main Results

• Scenario 1 (Baseline - Flood Extent + DTM): ResNet34 achieved the best performance with a Root Mean Squared Error (RMSE) of 0.226 meters, Mean Absolute Error (MAE) of 0.082 meters, Structural Similarity Index (SSIM) of 98.5%, and R-Squared of 93%. The Swin-Unet model showed significantly lower performance (RMSE = 0.518 meters, MAE = 0.207 meters, SSIM = 92%, R-Squared = 65%).
• Scenario 2 (Enhanced - Flood Extent + DTM + TWI, slope, curvature): ResNet18 emerged as the best-performing model, achieving an RMSE of 0.216 meters, MAE of 0.070 meters, SSIM of 98.8%, and R-Squared of 93.5%. The inclusion of terrain predictors improved ResNet18's Huber Loss by 28%, RMSE by 13%, and MAE by 21%. Swin-Unet also improved relatively (RMSE by 28.2%, MAE by 30.9%) but remained inferior to CNN models in absolute performance.
• Model Generalization: When the best-performing model (ResNet18 from Scenario 2) was applied to an unseen catchment, predictive accuracy declined, with RMSE increasing to 0.594 meters and MAE to 0.576 meters. This represented a 174.6% increase in RMSE and a 721.7% increase in MAE, attributed to limited and temporally misaligned ground truth data and spatial domain differences. Despite this, the model retained the ability to generate spatially consistent and plausible flood depth estimates.

Contributions

• Developed a comprehensive deep learning framework that fuses high-resolution remote sensing data with deep learning architectures, informed by hydrostatic principles and physically relevant variables, addressing a gap in existing literature.
• Demonstrated that lightweight ResNet-UNet architectures can accurately estimate flood depth under hydrostatic equilibrium conditions using remote sensing data.
• Quantified the significant improvement in mapping accuracy and spatial coherence achieved by incorporating Digital Elevation Model (DEM)-derived terrain features (slope, curvature, Topographic Wetness Index).
• Showcased the potential for feasible near-real-time urban flood-depth water mapping to support rapid disaster response, particularly in data-scarce regions, reducing reliance on in-situ measurements.

Funding

• NASA award 80NSSC23M0051
• NSF grant 2401942

Citation

@article{Blay2025Geospatial,
  author = {Blay, Jeffrey and Hashemi-Beni, Leila},
  title = {Geospatial and Deep Learning Approaches for Modeling Floodwater Depth in Urbanized Areas},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs18010060},
  url = {https://doi.org/10.3390/rs18010060}
}