A Deep U-Net Framework for Flood Hazard Mapping Using Hydraulic Simulations of the Wupper Catchment

1. Problem Statement

Flood events are increasing in frequency and severity due to climate change and land sealing. Traditional physics-based numerical models (PBNMs) for flood prediction are highly accurate but computationally expensive, making them unsuitable for real-time warning systems. While machine learning (ML) surrogate models offer a faster alternative, existing approaches face challenges:

• Generalization: Models often overfit to specific topographies and fail when applied to unseen catchments.
• Spatial Information: Many models ignore critical spatial data (e.g., Digital Elevation Models) or rely on simplified inputs.
• Inference Artifacts: Standard patch-based inference methods often introduce boundary artifacts that degrade prediction accuracy.
• Data Scarcity: Acquiring sufficient training data for new study areas is difficult.

The specific goal of this research is to develop a deep learning-based surrogate model that can rapidly predict maximum water levels across a grid based on river discharge and topography, serving as a computationally efficient alternative to hydraulic simulations for the Wupper catchment in Germany.

2. Methodology

A. Study Area and Data Generation

• Locations: Three distinct areas in the Wupper catchment (Germany) were selected: Beyenburg (BEY) for training, Kluserbrücke (KLU) for transfer learning, and Kolfurth (KOL) for zero-shot generalization testing.
• Input Data:
  • Topography: A 1m resolution Digital Elevation Model (DEM) (EPSG:25832).
  • Hydrology: Constant river discharge values (scalar).
  • Mask: A binary mask indicating valid grid points vs. no-data regions.
  • Output: Maximum water level maps generated by the 2D hydraulic simulation software HydroAS.
• Patch Strategy: Due to the large image sizes (up to ~10,000 x 10,000 pixels), a patch-based approach was adopted. Patches were randomly sampled using rejection sampling to ensure each contained at least one valid data point.
• Data Augmentation: To improve generalization and robustness, the training set was augmented with horizontal/vertical flips and rotations (90°, 180°, 270°). Input and target patches were augmented simultaneously to preserve physical relationships.

B. Model Architecture (Deep U-Net)

The authors utilized a U-Net architecture, a Convolutional Neural Network (CNN) designed for pixel-level segmentation tasks.

• Structure: Encoder-decoder with skip connections to preserve fine-grained spatial details (e.g., riverbanks).
• Optimization: Systematic ablation studies were conducted to determine the optimal Depth (number of encoder/decoder blocks) and Width (number of initial filters).
• Training: The model was trained using the Adam optimizer, RMSE loss, and early stopping. Normalization (min-max) was applied to inputs and targets.

C. Inference Strategies

To generate full-domain predictions from patch-based models, three strategies were tested to mitigate boundary artifacts:

1. No Overlap: Non-overlapping tiling.
2. Overlap: Sliding window with 50% stride; predictions in overlapping regions are averaged.
3. Center Crop: Only the central portion of each patch's prediction is used, discarding edge predictions to avoid boundary artifacts.

D. Evaluation Metrics

• RMSE (Root Mean Squared Error): To penalize large errors significantly.
• NSE (Nash-Sutcliffe Efficiency): To quantify predictive skill relative to the mean.

3. Key Contributions

1. Optimized U-Net Framework: The study identifies an optimal architecture configuration (Depth 4, Width 32 filters) that balances accuracy and computational complexity for flood mapping.
2. Inference Strategy Analysis: The paper demonstrates that the Center Crop inference strategy significantly outperforms standard non-overlapping or overlapping methods by eliminating boundary artifacts, reducing RMSE by nearly 25% compared to non-overlapping approaches.
3. Ablation Studies:
   • Patch Size: Larger patches (256x256) improved accuracy but increased training/inference time by ~3x.
   • Patch Amount: Increasing the number of patches per image improves performance up to a point of diminishing returns (around 400 patches), beyond which computational cost outweighs marginal accuracy gains.
   • Normalization: Target normalization was found to create a more robust, scale-invariant model, despite a negligible trade-off in raw RMSE on the source domain.
4. Generalization Analysis: The study rigorously tests "zero-shot" generalization on unseen catchments (KLU, KOL), revealing that models trained on a single catchment fail to generalize without multi-catchment training data.

4. Results

• Performance on Source Domain (BEY):
  • The final optimized model achieved an RMSE of 0.0227 m and an NSE of 0.9994.
  • Speed-up: The surrogate model generates a full flood map in 0.696 minutes, compared to 15 minutes for the traditional hydraulic model. This represents a ~21.5x speed-up.
• Inference Strategy:
  • Center Crop: Best performance (RMSE 0.037 m).
  • Overlap: Moderate performance (RMSE 0.0443 m).
  • No Overlap: Worst performance (RMSE 0.0487 m) due to visible grid artifacts.
• Generalization (Zero-Shot):
  • When applied to unseen catchments (KLU and KOL) without retraining, the model performed poorly (NSE < -1.0), indicating it memorized specific topographical features of BEY rather than learning universal hydraulic laws.
• Error Analysis:
  • Errors were generally low but clustered around complex man-made structures (dams, bridges).
  • A key finding was a discrepancy in data preprocessing: the hydraulic model's DEM had bridges removed to guide flow, while the surrogate model's training DEM included them, leading to prediction errors near these structures.

5. Significance and Limitations

Significance:

• The research validates that deep learning surrogate models can serve as a viable, high-speed alternative to physics-based models for real-time flood warning systems.
• It provides a reproducible methodology for handling large-scale spatial data through patching and specific inference strategies (Center Crop) to ensure high fidelity.
• The 21x speed-up makes it feasible to integrate flood mapping into real-time warning systems like the "Bergisches Hochwassermeldesystem 4.0."

Limitations:

• Lack of Real-World Validation: The model was trained and tested solely on simulated data, not real-world observational flood data.
• Generalization Failure: The model failed to generalize to unseen topographies (KLU, KOL) in a zero-shot setting, highlighting the need for multi-catchment training datasets.
• Input Simplification: The model used constant discharge and static DEMs, ignoring dynamic variables like soil saturation, slope, or time-varying hydrographs.
• Preprocessing Discrepancies: Inconsistencies in how structures (bridges) were handled between the physics-based and data-driven pipelines caused localized errors.

Future Work:
The authors suggest validating models with real sensor data, incorporating dynamic inputs (hydrographs, precipitation), and developing strictly consistent data pipelines to handle structural features uniformly across physics and data-driven models.