Using Image-Based Deep Learning to Identify Levees from Elevation Data for National-Scale Flood Modeling in the United States

Abstract:

National-scale flood hazard maps are essential tools for assessing property risk and the financial impacts of flooding. The provision of undefended hazard maps with separate defense information, rather than the inclusion of defenses in the modeling process, provides risk practitioners with the greatest flexibility when assessing flood risk. However, there is a global shortage of information on the locations, standard of protection and state of repair of defenses. For example, in the United States it is estimated that there are approximately 100,000 miles (160,000 kilometres) of levees, but the location of many of these is not mapped in national-scale datasets (USACE, 2018). We present an automated approach for the large-scale identification of levees using deep learning techniques.

The full or partial representation of levees in the Digital Elevation Model (DEM), used in the hydraulic modeling process, reroutes out-of-channel flow producing unrealistic flooding in undefended flood maps. To generate undefended flood maps these levees need to be entirely removed from the DEM, which requires knowledge of their locations. Without comprehensive levee datasets, an alternative method to identify levees in DEMs on a large scale is required.

The use of deep learning techniques to recognise objects in images is fast developing. DEMs and other related datasets can be represented in a similar format to images. We have trained an image segmentation model using the U-Net Convolutional Neural Network architecture to identify levees in DEMs. We have used the resultant levee dataset to generate undefended flood maps for the whole of the contiguous United States. We present details of the methodology, including how we trained the model, and the challenges faced when applying the model to the varying terrains across the US.