Machine Learning Classification of Flood Waters from Hurricanes Harvey and Florence as Captured by Synthetic Aperture Radar and Optical Remote Sensing

Major hurricane events impacting the United States result in significant resources directed by federal agencies towards response to those affected. To assist with determining the severity and scope of flood-affected areas, many partners in academia, state emergency managers, the private sector, and federal agencies have developed methods to apply optical and synthetic aperture radar (SAR) remote sensing to the mapping of flood water extent. For example, activations of the International Charter: Space and Major Disasters brings together complimentary, international sources of optical and SAR imaging. As weather conditions improve, NOAA and the Civil Air Patrol fly missions to collect true color photography at high spatial resolution. Routine optical imaging from NASA, NOAA, and international satellites provide water mapping from true color, near-infrared, and thermal data, complimented by others who map impacts through passive microwave. Unique to Hurricane Harvey of 2017 and Hurricane Florence of 2018, NASA and the Jet Propulsion Laboratory (JPL) flew missions of the Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) sensor aboard a Gulfstream-III aircraft, collecting numerous swaths over the Houston metropolitan area (2017) and record-setting river stages across North and South Carolina (2018). The resulting L-band, polarized collections of SAR imaging allow for high spatial resolution mapping of water extent, and a unique advantage over optical imaging in the ability to better penetrate deep vegetation canopies of the forested Carolinas and vegetated regions of the suburban Houston area.

In this presentation, we use NOAA aerial and complimentary commercial satellite imaging (Planet/DigitalGlobe) to perform supervised classification of UAVSAR flood scenes using a Random Forest model focused on the mapping of open water and flood water otherwise obscured by vegetation, based on flight tracks acquired during Hurricanes Harvey and Florence. We also develop classification of flood water from commercial satellite imaging to complement UAVSAR and other international SAR and optical imaging collections. Output from the Random Forest classifier is examined for value and skill in both deterministic classifications of scenes as well as probability information for various classes, particularly those focused on open and obscured flood waters towards beneficial products supporting analysts involved in response efforts.