A Database of Flood Maps using high-resolution Airborne Imagery and Machine Learning Models

Optical Satellite imagery commonly suffers from the presence of cloud cover during flood events; Radar Satellites are disadvantaged from water look-alike conditions where the ground surface interacts with the incoming radar signal as if it were water; Regardless of modality of satellite, more importantly, satellite overpasses during a flood are chance occurrences where the capture of the maximum extent is a fortuitous incident. Low-altitude aerial remote sensing, on the other hand, can be used to survey the extent of flooding at the peak or soon after it has occurred, with a good measure of reliability. Opportune scheduling of these reconnaissance flights not only capture floods at ultra-high resolution, but also allows for seamless geographical coverage unhindered by cloud cover.

The National Oceanic and Atmospheric Administration (NOAA) Emergency Response Imagery is very high resolution (50 cm Ground Sampling Distance between pixels) airborne imagery acquired by the Remote Sensing Division of the National Geodetic Survey (NGS) during major flood events in the United States to support NOAA’s homeland security and emergency response requirements.

In this work, we evaluated the performance of four different machine learning models (Gradient Boosting, Random Forest, Support Vetor Machine, Convolutional Neural Network) on the ability to classify floods from raw aerial imagery. The classifier with the highest classification accuracy metrics - depending on geographical and hydrological setting – was used to produce flood inundation extent maps for 30 major flood events.

We demonstrate the utility of these high-fidelity flood maps via two use-cases: both synergistic studies to this work. 1) As benchmarks for validation of hydrodynamic model results: Historic flooding occurred on the Neuse River in North Carolina in the United States triggered by Hurricane Matthew in 2016. Several hydrodynamic models were deployed to simulate flood dynamics with an end goal of understanding flood susceptibility in the Neuse basin under changing climate conditions. The aerial imagery-based flood maps were used as benchmarks for model validation. 2) Enhancing the versatility of FIMPEF: Flood Inundation Mapping Predictions Evaluation Framework (FIMPEF) is an open-source, cloud-based geospatial venture by the University of Alabama that calculates accuracy statistics between benchmark and modeled flood extents. Integration of aerial imagery, in addition to the satellite-based benchmarks that FIMPEF was ingesting so far, has vastly enhanced its robustness and user-demand. Free access (no account/login credentials needed) to these high-quality flood maps is granted through the United Sates Flood Inundation Map Repository (USFIMR), an online geospatial warehouse that provides high-resolution inundation extent maps of past U.S. flood events.