Intercomparison of Satellite Remote Sensing-Based Flood Inundation Mapping Techniques

The objective of this study was to determine the accuracy of five different digital image processing techniques to map flood inundation extent with Landsat 8-Operational Land Imager satellite imagery. The May 2016 flooding event in the Hempstead region of the Brazos River, Texas is used as a case study for this first comprehensive comparison of classification techniques of its kind. Five flood water classification techniques (i.e., supervised classification, unsupervised classification, delta-cue change detection, Normalized Difference Water Index [NDWI], modified NDWI [MNDWI]) were implemented to characterize flooded regions. To identify flood water obscured by cloud cover, a digital elevation model (DEM)-based approach was employed. Classified floods were compared using an Advanced Fitness Index to a reference flood map created based on manual digitization, as well as other data sources, using the same satellite image. Supervised classification yielded the highest accuracy of 86.4%, while unsupervised, MNDWI, and NDWI closely followed at 79.6%, 77.3%, and 77.1%, respectively. Delta-cue change detection yielded the lowest accuracy with 70.1%. Thus, supervised classification is recommended for flood water classification and inundation map generation under these settings. The DEM-based approach used to identify cloud-obscured flood water pixels was found reliable and easy to apply. It is therefore recommended for regions with relatively flat topography.