Inundation2Depth: A multi-source dataset for floodwater depth estimation in urban areas

Area of Science:

• Geospatial Artificial Intelligence (GeoAI)
• Remote Sensing
• Hydrology

Background:

• Urban floods pose significant risks, but operational mapping often lacks critical flood depth data.
• Developing accurate flood depth models is hindered by the scarcity of large, georeferenced, and well-labeled datasets.
• Deep learning offers advanced solutions but requires substantial, high-quality training data.

Purpose of the Study:

• To introduce the Inundation2Depth dataset, a novel resource for flood depth estimation.
• To facilitate the development and evaluation of GeoAI models for urban flood severity assessment.
• To lower data barriers and promote comparability in flood-related GeoAI research.

Main Methods:

• Paired inundation extent and depth labels derived from aerial imagery and LiDAR (Light Detection and Ranging)-based Digital Terrain Models (DTMs).
• Data collected from 12 flood-affected areas in North and South Carolina (2016-2018) covering diverse environmental conditions.
• Multi-sensor remote sensing data (Optical and LiDAR) preprocessed for spatial consistency and provided as raster tiles for machine learning integration.

Main Results:

• The Inundation2Depth dataset comprises 5925 overlapping tiles across 24,649.88 acres.
• Data includes raw and normalized versions, suitable for direct integration into machine/deep learning pipelines.
• Dataset validation performed using hydrodynamic modeling with the HEC-RAS Rain-on-Grid tool.

Conclusions:

• The Inundation2Depth dataset addresses the critical need for georeferenced flood depth data.
• Its standardized format and spatial diversity enhance its value for flood detection, segmentation, and damage assessment models.
• This resource is expected to advance GeoAI research in urban flood severity and improve flood risk management strategies.