Towards global flood mapping with machine learning based on the Harmonized Landsat-Sentinel 2 data

Abstract

Floods are the largest natural hazard in terms of life loss and economic damage, regardless of their cause. In the United States alone, floods cause billions of dollars in property damage, with an estimate exceeding $78 billion due to fluvial floods in any given year. Effective and immediate disaster response management can reduce the impact of floods but it requires near real-time information on flood occurrence. To best allocate limited resources and prioritize response actions during hazardous floods, emergency responders need near real-time information on flood-water extent, typically derived based on Earth Observation (EO) data. Satellite remote sensing offers the only means of monitoring and quantifying flooding extent dynamics and the availability of public domain, systematically acquired satellite data archives, together with improvements in algorithms and available computing power have led to huge leaps in the recent years in mapping surface water dynamics and flooding. A large proportion of the prior work has relied on optical data, including MODIS and Landsat thus trading high temporal resolution with daily maps in the case of MODIS or higher spatial resolution but coarser temporal resolution in the case of Landsat. However, the recent availability of NASA’s Harmonized Landsat/Sentinel-2 (HLS, https://hls.gsfc.nasa.gov/) Surface Reflectance Product, a seamless data set combining Landsat 8 and Sentinel 2 observations, is promising in detecting floods at Landsat resolution and 3-day interval. New work in a dryland basin that experiences ephemeral floods showed that large short-lived flooding events were detected only by HLS (the combined dataset) but have been entirely missed by Landsat 8. Here we picked major flood events globally labeled with collocated Harmonized-Sentinel-2 data and applied machine learning models for flood detection. The most important features for flood detection included the SWIR bands, the automated water extraction indices, and vegetation indices. Future work will integrate Sentinel 1 radar data collocated to the Harmonized-Sentinel-2 data for improved detection of floods during cloudy conditions. This work also highlights the importance of existing harmonized data products such as HLS.

Mirela G. Tulbure

Professor

I am an Associate Professor with the Center for Geospatial Analytics at North Carolina State University (NCSU).

Mark Broich

Postdoctoral Research Fellow

Played a central research role on the Geospatial Analysis for Environmental Change team and had a key role in building the GAEC lab.

Mollie D. Gaines

Postdoctoral Research Scholar

I am a Postdoctoral Research Scholar in the Department of Forestry and Environmental Resources at North Carolina State University.