FloodSimBench: A High-Resolution Physical Flood Inundation Benchmark for AI

FloodSimBench is a standardized, high-resolution (1m) benchmark dataset designed for urban flood inundation modeling. It bridges the gap between complex physical hydraulic modeling and deep learning foundation models (e.g., Transformers, CNNs).

1. Overview

• Objective: Provide a high-fidelity dataset for training and evaluating AI surrogate models for flood risk assessment and real-time forecasting.
• Domain: Diverse urban topographies across 10 major US cities (coastal, inland, flat, and hilly).
• Core Value: Standardized inputs (DEM, Rainfall) and outputs (Water Depth) at 1m resolution to enable comparable research in the AI community.

2. Dataset Specifications

• Spatial Resolution: 1 meter.
• Domain Size: Approximately 4km x 4km per city tile.
• File Format: GeoTIFF (.tif).
• Simulated Event:
  • Forcing: 1 hour of rainfall (00H-01H) at varying frequencies (10yr, 25yr, 50yr, 100yr), followed by 5 hours of recession (no rain).
  • Total Duration: 6 hours.
  • Temporal Resolution: 5-minute intervals (72 frames per scenario).

3. Metadata & Scenarios

Rainfall intensities based on return periods for the covered cities:

City ID	Description	100-yr (mm)	50-yr (mm)	25-yr (mm)	10-yr (mm)
HOU	Houston	129	110	98	82
AUS	Austin	121	106	93	76
DAL	Dallas	94	85	76	65
OKC	Oklahoma City	101	89	78	64
LA	Los Angeles	41	36	31	25
SF	San Francisco	33	29	26	22
NYC	New York City	73	65	58	48
ATL	Atlanta	83	74	65	54
ORL	Orlando	100	93	85	74
MIA	Miami	133	118	104	86

4. Benchmark Tasks

Task A: Static Risk Assessment (Segmentation)

Goal: Predict the maximum inundation extent and severity over the 6-hour event.

• Input: Static DEM, Slope, and Total Rainfall Amount.
• Output: 5-class Semantic Segmentation Map:
  0. No Flood (0 - 0.1m)
  1. Nuisance (0.1 - 0.2m)
  2. Minor (0.2 - 0.3m)
  3. Moderate (0.3 - 0.5m)
  4. Major (> 0.5m)

Task B: Spatiotemporal Forecasting (Recursive)

Goal: Real-time recursive flood forecasting.

• Input: Static DEM and Dynamic Flood Depth sequence ($t-11 \dots t$, previous 1 hour).
• Forcing: Rainfall intensity at $t$ and future steps.
• Output: Predicted flood depth at future time steps (e.g., $t+1 \dots t+12$).

5. Dataset Structure

.
├── 6hr_max/               # Maximum depth maps for Task A
│   └── {CityID}_{Rain}mm_MaxDepth.tif
├── {CityID}{TileID}/      # Full time-series for Task B
│   ├── {CityID}_DEM.tif   # Static Digital Elevation Model
│   └── {CityID}_{Rain}mm_WaterDepth_{Time}.tif # Temporal depth frames
└── cities_rainfall.json   # Metadata for rainfall intensities

6. Data Splitting Strategy

To ensure robustness and test generalization to unseen topographies, we split by City:

• Train: HOU, DAL, OKC, NYC, ATL, ORL
• Validation: AUS, MIA
• Test: LA, SF

7. Citation

If you use this dataset in your research, please cite:

(Citation info pending publication)