Authors: Amobichukwu C. Amanambu—University of Alabama
Title: U-NETwet : A Deep Learning Approach for Predicting River Depth and Floodplain Inundation
Abstract: This study presents an innovative deep-learning framework used to model river depth and floodplain inundation. A high-resolution river bathymetry and surface (1-m resolution digital surface and river elevation model—DSRE) model was created by combining a multi-beam hydrographic survey and LiDAR point cloud. The combined data was converted to a channel-centered spatial reference system (s, n) from a geographical coordinate system (x, y) through forward and inverse transformation technique. The generated DSRE served as an input into a 2D HEC-RAS model for creating a river depth and floodplain inundation surface. A convolutional neural network (CNN) architecture for image segmentation, U-NETwet—modified, was used to train a model that used flow patterns and topography to predict water depths across the river and floodplain. The configuration of the U-NETwet (128 × 128 × 2) takes two channels (bands) of stream flow and topography with a single output that measures depth and width. The results reveal that the U-NETwet can efficiently identify and predict the river shape, depth, and floodplain wetted areas. A relative elevation water surface model (REM) was used to approximate flood inundation patterns from the HEC-RAS model with a maximum difference of 1.7 m. The result of the U-NETwet and the 2D hydrological model was compared using mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R22). The evaluation metrics show a significant agreement with the two models with an = 0.0011—0.0042, = 0.00097—0.004, and =0.90—0.98. The error difference between the two models is < 0.5 m. Both models reveal the same direction with increasing discharge. The result suggests that deep learning can serve as a fast and automatic way to model inundation accurately.